YT YouTube2Text

YouTube2Text → Transcripts → AI for Safer Roads Final Pitch Event

AI for Safer Roads Final Pitch Event — Transcript

by ADB Digital Sandbox · 18,434 words · 2,795 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:01Good morning and good afternoon and a
0:03very warm welcome to the final pitch
0:06event of the AI for safer roads safer
0:09speeds innovation challenge 2026.
0:12My name is Tia Carter. I'm an open
0:14innovation coordinator at the
0:16information technology department at the
0:18Asian Development Bank and I'll be your
0:20host this afternoon.
0:23Let me start by introducing the partners
0:26behind this challenge. This innovation
0:28challenge has been brought to you by ADB
0:31transport sector cluster and the
0:33information technology department in
0:36collaboration with the Asia Pacific road
0:39safety observatory and world bank group
0:43world bank development group and
0:45international telecommunications union
0:48as well as AI for good and
0:51with support from Japan fund for
0:53prosperous and resilient Asia and the
0:56Pacific and the high technology fund.
1:00Earlier this year, ADB and its partners
1:02asked the innovators everywhere a
1:04question. Could AI show safety experts
1:09and governments where their road speeds
1:12limits no longer fit the roads and have
1:15the speed set according to the current
1:17conditions before accidents even take
1:19place. 46 teams from 15 countries took
1:23it on. Seven were selected for an
1:25intense one and a half month building
1:27and refinement phase and four were
1:30selected to be here today to pitch their
1:32solutions. We have an exciting program
1:35today. The ADB leadership will open the
1:38program in just a few minutes after
1:41which our colleague Pritiki Kam will
1:43provide context to the road speed
1:45challenges in Asia and the Pacific.
1:48Then we will introduce today's stars,
1:51our finalist teams and the members of
1:53the jury before kicking off to the most
1:56exciting part of today, the four
1:58finalist pitches.
2:00Each team will have five minutes to make
2:02their case followed by five minutes of
2:05questioning by the jury. After we finish
2:08with the pitches, the jury will move to
2:11a separate meeting to deliberate while
2:13you audience can vote for the audience
2:16choice award. So listen closely to all
2:20four pitches as you will be asked to
2:22vote for your favorite. After this we
2:25will have a panel discussion focused on
2:27the implementation and impact of
2:29transport sector innovation moderated by
2:31greater than CEO Yana.
2:35And after the judges return, we will
2:37announce the selected solutions and take
2:39a moment to celebrate their
2:41achievements. Before closing our event
2:43for today,
2:46the team taking the first place will
2:49have the opportunity to pilot and
2:51further develop their solution together
2:53with ADP with a contract up to 20,000
2:56USDs. Furthermore, ITU may provide cash
2:59prices to the first and second place.
3:04Before we move on, a quick housekeeping
3:07notes for the presenters. Please
3:10remember to keep your mics muted when
3:12not speaking.
3:14Secondly, um
3:17audience, uh you are invited to use the
3:19Q&A box for questions to the teams. Just
3:22clearly indicate who the question is
3:24for. You can also participate in the
3:27conversation on the chat. And lastly,
3:30the session is uh being recorded and
3:33will be shared on the ADB challenges
3:35website afterwards.
3:38Now, without further ado, to open
3:40proceedings, it is my pleasure to
3:43welcome Mr. Hideyaki Iwasaki, director
3:46general of sector department one at the
3:49Asian Development Bank. Thank you, Mr.
3:52Iwasaki, for being here today. The floor
3:54is yours.
3:56>> Thank you. Um yeah
3:59uh good morning and good afternoon and
4:02good evening. Yeah I'm I'm very happy to
4:06uh open this uh uh challenge um
4:12which um is intended as TA mentioned to
4:18fully utilize the potential of AI in
4:22addressing the perennial issue of road
4:24safety. Um
4:28road crashes still claim
4:31uh a lot of lives and also uh it is a
4:35huge uh socioeconomic burden of many of
4:40the countries in the world and
4:42especially in uh the countries
4:46uh in Asia and the Pacific
4:49and uh addressing this challenge uh we
4:53have been investing
4:56in uh infrastructure and also investing
5:01in the people uh to change their
5:06behavior
5:07and also um the the overall ecosystem.
5:13uh but um
5:17now we have uh the access to the
5:20technology especially AI in the recent
5:23past uh uh its capability has really uh
5:28increased
5:30and combining
5:32AI and GIS and other technologies
5:38uh we should be able to uh address this
5:42challenge.
5:43in a more uh effective and efficient
5:47manner. Um this is the genesis of uh
5:52this challenge. Um
5:56ADB has been uh providing
6:01policy dialogue,
6:03technical assistance
6:06and
6:09financial resources.
6:11uh both to create uh the ecosystem
6:16and also uh create the necessary
6:19infrastructure.
6:20Um but we are also increasingly uh
6:24increasing our digital capabilities.
6:27Uh my colleague uh from u the uh IT
6:32department Stephanie is joining and
6:34she's speaking soon after me. Um and we
6:39uh with increased capability we want to
6:41bring innovations.
6:45Um this uh challenge is possible uh by
6:50working with uh so many uh partners. Um
6:55I'm not going to name each of the
6:57partners but I really appreciate uh
7:00their contributions
7:02uh to this initiative and we would like
7:07to uh
7:10go beyond this challenge
7:13u to uh explore further opportunities
7:18to apply new technologies
7:22uh to uh addressing challenges. Road
7:28safety is
7:30one of the main challenges we face but
7:34there are many other challenges and we
7:36would like to utilize innovations
7:39uh in uh those efforts.
7:43um uh the the finalist. I'm really
7:48looking forward to listening to uh your
7:51pitches and as TA mentioned uh the
7:56winner will be given the opportunity to
8:00put the pilot ideas into practice
8:05uh with ADB's facilitations. So I'm
8:08really looking forward to uh seeing the
8:12you know the innovation put into
8:14practice. Um thank you again for um
8:20joining this uh you know challenge and
8:24uh I I wish you all the very best for
8:26your pitches. Thank you.
8:30>> Thank you very much u Mr. Usaki. Thank
8:33you for joining us today. Next, I would
8:36now like to welcome Miss Stephanie Hung,
8:39the chief information officer and
8:41director general of the information
8:43technology department at the Asian
8:45Development Bank. Thank you Stephanie
8:47for joining us today. You may take the
8:50floor.
8:50>> Yeah. Uh thank you uh Taja and also
8:53thank you Hideiaki for setting up why
8:56role safety matters for our developing
8:59member country and how we could actually
9:02utilize AI to innovate um those effort.
9:06So good day everyone and uh thank you
9:08for having me today. Um I would like to
9:11pick up where um Hideyakis left off is
9:15talk about what it takes for an
9:17institutions like uh ADB to actually put
9:21AI to work. So across our region um that
9:25is no longer a um hypothetical question
9:30whether governments in Asia uh you know
9:33uh should deploy. Um uh the question
9:36actually have shifted. It is no longer
9:38whether AI can contribute to issues such
9:42as road safety. It is whether
9:44institution like ours can actually help
9:48our members to member country to adopt
9:51it safely and responsibly in a way that
9:56actually um uh hold up long after the
10:00pilot uh end. So that is where uh today
10:04I wanted to uh make sure that you know
10:07uh responsible AI transformations are
10:09successful. Uh I would like to look at
10:12three aspect. The first is really about
10:15technology and infrastructure must be
10:18make practical for developing member
10:21country. So what it mean is really
10:23localized model, modular architecture,
10:27cost effective deployment, not the most
10:31powerful uh system or model available
10:34but one that a country can run and
10:37afford. Second, it is really about uh
10:41governance and trust. So uh within ADB
10:46we actually has actually um a landmark
10:50uh responsible AI framework which is we
10:53are now operationalizing it. So it is
10:56guided by eight principle.
10:59People and planet first
11:03safety and reliability,
11:06explanability,
11:09transparency,
11:11privacy and security,
11:14data integrity,
11:16fairness and inclusion, and the most
11:19important accountability. A human center
11:22AI must be deployed responsibility
11:26if it is to earn public trust. So in
11:31this scenario when we talking about road
11:33safety uh reliability safety is utmost
11:37but at the same time we also need to
11:40ensure that we assess the people harm
11:45environmental harm and financial harm
11:48and we actually have a risk mitigation
11:51approach towards that. What matters is
11:54that we have not left those as a
11:57principle. What we did is that we
12:00actually translate all of them into a
12:03technical verifiable automated control
12:06and we actually also launch our
12:08responsible AI risk assessment platform
12:12with a life cycle management covering
12:14every AI use cases and we also set up AI
12:19subcommittee which is composed of all
12:22the risk professional uh from our um
12:26general council to our anti-corrupt
12:28ruptions to our risk management team to
12:31our department of communications and
12:34knowledge and uh HR these are the people
12:37who then actually make the assessment of
12:40the uh risk that we are dealing with.
12:43Then at the same time we also have uh
12:45you know the ADB's uh you know uh
12:47sandbox for us to actually have a safe
12:50facility for us to actually experiment
12:54uh the business model the technology and
12:57then from there we can then actually
12:59move that into minimum viable product to
13:02subsequently uh digital product and uh
13:06the last one it is really on people. So
13:10um b uh because the ultimate value of uh
13:13AI would not be measured by the size of
13:16our data center or the number of uh
13:20model parameters that uh we can process.
13:23It is actually measured by whether it
13:26improve life whether it makes public
13:29services more accessible more
13:32transparent and more effective. So
13:34that's bring me today this challenge
13:37would put these uh all three foundation
13:40in practice and what interests me about
13:43the solution that uh you will see is not
13:45the sophistication of the model but they
13:49are built to be used uh the work from
13:53the data that already exists designed
13:56for planners rather than data scientists
14:00and are able to show their reasoning. So
14:02when a me recommen recommendation that
14:06touches something as a consequential as
14:08a speed limit uh a planner needs to see
14:11why a role was flat and be able to
14:15defend it and that is explanability and
14:18accountability in practice rather than
14:20in principle. So our goal today is not
14:23only to select the winner but also to
14:26show what this tool can do to discuss
14:29what it takes to move from pilot to real
14:32use at scale and to start building a
14:35community around datadriven uh you know
14:38speed management. So finally uh let me
14:41take this moment thank you all our
14:43partner and jurors who has given their
14:46time and genuine expertise to this the
14:49organizing team who brought it together
14:51and to the team that who are presenting
14:54today congratulations on getting here
14:57and I'm looking forward to seeing what
14:59you have built. Uh thank you and enjoy
15:02the event.
15:04Thank you both for those opening remarks
15:07and for showing ADB's commitment to
15:10embrace innovation for good and to
15:12address this very important development
15:15challenge.
15:17Now to set the scene, I would like to
15:19hand over to Pritik Kam, senior
15:22transport specialist and road safety
15:25lead at ADB,
15:27the person who has been championing this
15:30challenge from the beginning and
15:32bringing everyone together around this
15:34issue. She will give you context on why
15:37this challenge on safer speed was
15:39launched and why it was an important
15:41issue. Pretty the floor is yours.
15:47Great. Thank you, Taj. Give me just a
15:49moment to pull up my screen. [snorts]
15:52Are you able to see it?
15:55>> Yes, we are.
15:57>> Okay. Excellent. [clears throat]
15:58Good afternoon everyone. Um, my name is
16:00Pratik Gautam and the work that we're
16:03doing here today contributes to a
16:06broader vision that we have for the Asia
16:09and the Pacific, which is to build safe
16:11and sustainable mobility. And what do we
16:13mean by that? We mean transport systems
16:16that work for the people, that work for
16:18the economy, and that work for the
16:20environment. It means transport systems
16:22where people can walk safely and
16:25efficiently without any risk of death or
16:28severe injuries. As you've heard in the
16:31opening remarks, unfortunately, much of
16:33the Asian, the Pacific has a long
16:35journey towards that vision. One person
16:38dies every minute on the road in the
16:40region. And this region has a larger
16:44share of road traffic fatalities and
16:46injuries than any other region in the
16:48world. It has a huge economic cost. But
16:51it also road safety has a business case.
16:55Road investments that decrease crashes
16:58actually have a significantly higher
17:01return than those that increase crashes.
17:05And historically, part of the challenge
17:07that we've had uh is that the way that
17:10we've addressed road safety has been
17:12quite reactive. We wait for a crash to
17:14happen. We analyze the data uh and then
17:18we retrofit our solutions. But
17:21technology, AI, and innovation offers us
17:24the opportunity to be more proactive and
17:28predictive in the way that we address
17:30road safety through identifying risks
17:33earlier, uh through integrating
17:35real-time data and through making
17:38smarter investments before a crash
17:40actually ever occurs.
17:43But technology won't go very far without
17:47a framework in place to underpin it. And
17:50for us in road safety, the gold star is
17:53the safe systems approach. And the safe
17:55systems approach recognizes that we as
17:58human beings, we will make mistakes, but
18:00no mistake should result in a death
18:04sentence or a severe lifelong
18:07disability.
18:08Uh it also recognizes that we all have a
18:11responsibility and the ability to design
18:15a transport system that when a crash
18:17occurs that the kinetic energy and
18:21forces that are exchanged within the
18:23system never reach a threshold
18:26that would cause severe consequences to
18:29the human body. And we know we can do
18:31that this by the way that we build our
18:33vehicles, by the way we build our
18:36infrastructure and the way that we
18:38manage speed. And that's why this
18:40innovation challenge focused on speed is
18:43so important because speed is the
18:45leading risk factor in the likelihood
18:47and severity of a crash.
18:51ADB, as you've heard from our DEG,
18:53supports road safety in many ways
18:55through financing, technical assistance
18:58or capacity development. Uh we also just
19:02this year finalized a pilot project
19:05where we challenged ourselves to use new
19:07technology to collect this more
19:09predictive uh data and and we call the
19:12safety performance indicators and this
19:15supports the UN global road safety
19:17performance targets. And we looked at
19:19three indicators. Um this was first um
19:23the part of the road network carrying
19:2575% of the traffic. So this looks at
19:29most you know which roads are the most
19:31used. So that's linked to exposure and
19:33risk. Second, we looked at the
19:36proportion of traffic traveling over the
19:38speed limit. We know this is important
19:40because a 1% increase in speed leads to
19:44a 4%
19:46increase in the risk of a fatal crash.
19:50And third, we looked at the proportion
19:52of two wheelers wearing helmets. We know
19:54that wearing a quality helmet can
19:56significantly decrease the risk of a
19:59brain injury and fatality of a
20:01motorcyclist. So the data behind this
20:04pilot project is what is the foundation
20:08of the data that the teams are using for
20:11this innovation challenge. So I'll share
20:13some of the results of our pilot project
20:16we we launched this pilot in Maharashtra
20:19India and the whole country of of
20:22Thailand. Uh and for the first indicator
20:25of exposure and road use, we used
20:28open-source road network data combined
20:30with GPS data and also GPS probe data
20:34from TomTom. And we looked at the whole
20:36year of 2024. So we're talking about
20:39billions of data points. And so if you
20:41were to imagine how much that time that
20:44would take for us manually to collect
20:47this data, you can imagine how resource
20:50intensive it is and what technology has
20:53to offer us in terms of scale. And what
20:56we found um for this particular
20:58indicator is that for places like
21:00Thailand and Maharashtra, India, only 12
21:04to 33% of the entire road network
21:07carried the most traffic. And why this
21:10is important is because when we have
21:12limited resources, we can hone in on the
21:15road corridors and the roads that have
21:19the most exposure and risk for
21:21prioritizing maintenance, design,
21:23upgrades, and enforcement.
21:26The second um indicator we looked at was
21:30speed. Uh and we looked at measuring
21:33vehicle speeds and the compliance to
21:35speed limits. And for this we used GPS
21:38data and we used TomTom data for speed
21:41limits. And we only included free flow
21:43traffic and avoided intersections and
21:46congestion. And all of this as you see
21:48on your screen is shown on a dashboard
21:50so that project teams, governments,
21:52policy makers can zoom in uh on specific
21:56corridors and see risk patterns and link
21:59that directly with infrastructure
22:01design, with policy decisions and
22:04enforcement.
22:06The last indicator we looked at was
22:08helmet wearing. And we used AI and
22:10computer vision to measure helmet use
22:12amongst motorcyclists, riders, and
22:15passengers. And for this, we used
22:17millions of crowdsource uh imagery from
22:20a pillar. And we trained an AI to detect
22:23the helmets, the helmet use. uh and that
22:27data was also put on that same dashboard
22:30where governments could and teams could
22:33zoom into specific cities and provinces
22:36and see helmet wearing rates. So this
22:39indicators as as you can see can be used
22:42before a project to inform targeting
22:45design during implementation to see you
22:48know whether what we're doing is
22:50actually reducing risk and also to see
22:52before and after the road safety
22:55performance. It of course can also
22:57inform policy decisions and enforcement.
23:00So our participants and our finalists
23:03had access to all of this data and we
23:05really asked them to hone in to that
23:07speed component of the safe systems
23:09looking at do the speed limits that
23:12currently stand actually reflect the
23:15function and use of the road. uh and so
23:19they are challenged to make an
23:20analytical model a speed safety s uh
23:24score and geospatial visualization and
23:28any viable solution that we have today
23:30would be integrated into ADB's platform
23:33used in our operations and made
23:35available to our developing member
23:38countries. So that's a little bit of a
23:40background of why we're here today and
23:42what we're aiming to achieve. Thank you.
23:45Back over to you Taja. Thank you,
23:48Pretty, for helping us know what is at
23:50stake and why this challenge is so
23:52timely and important.
23:56Now, let's meet the stars of today, our
23:59finalists as well as our esteemed
24:02members of the jury who will be
24:04assessing the team pitches today.
24:07Four teams are with us today presenting
24:10in this order. First, we have Neural
24:13City. Neural City is a Indian civic tech
24:16company whose platform turns near
24:18realtime data city data into analytics
24:20that help administrators decide where to
24:23act. First team members include Brites
24:26Vidi Akan Shupojid Sakraorti and Milan
24:32Gupta. Next we have Speech by Hanoi Wong
24:36an independent entrant working in AI
24:39policy analysis and product design.
24:42She's based in Washington DC, USA.
24:46Next we have team Kodi by IDN Nippon Koi
24:50Business Partners Singapore branch. Team
24:53Kodi is passionate in using data
24:56science, AI and machine learning to
24:58solve business and societal problems and
25:01team members include Manbon or Andy
25:04Wong, Guyan, Shan Lin, Muhammad Igandar
25:08and Yukina.
25:10And last but not least, team 5S formed
25:13for this competition by Georgia based
25:16road safety professional Kentaro Eio and
25:19India based professional engineer Vivbec
25:22Gupta. Kentaro operates 110 Levid LLC
25:27and Vivbec represents Deep Traffic
25:29Private Limited Company. Together they
25:32provide a tool that shows road agencies
25:35with speed limits to act first. Great to
25:38have you all. Thank you for being here
25:41with us today.
25:43Now let's move on to the jury. Assessing
25:45the finalist today are
25:49um Tim Manayar, program director at
25:53TomTom. She's a program leader and risk
25:56mitigation specialist with 25 years
25:59experience in mapping and is passionate
26:01about diversity and people engagement.
26:05Dimma um I would like to ask you what
26:08are the two major considerations you
26:11will be listening in for in today's
26:13speeches.
26:15>> Yeah thank you so much Nisa. So um hello
26:19everyone. Um to be very honest I'm very
26:21thrilled uh and excited to be part of
26:24the jury today as we look at how these
26:26teams are tackling one of our most
26:28critical urban challenges that's road
26:30safety.
26:32So I will be tuning into three key
26:34pillars. Uh the first is the data
26:36integrity and the risk modeling. Uh and
26:39uh which means like how accurately does
26:41the AI isolate where speeds actually put
26:44road users at risk. That's the first
26:46one. The second is the actionable
26:49usability. Um I just want you to check
26:51can a city planner or traffic engineer
26:54open this interactive map and instantly
26:56know where to intervene first. That's my
26:59second pillar of thought. And the third
27:01is the scalability. uh how seamlessly
27:04can local authorities deploy and
27:06maintain this across a real world
27:08network. That's my third pillar to
27:10check. And uh what I'm most looking
27:12forward today is seeing how these teams
27:15transform complex geospacecial and
27:17telemetry data into intuitive
27:19life-saving decision tool. So I'm not
27:22here just to judge the coding. Uh I'm
27:24also interested to check uh of course
27:27the solution that can genuinely shape
27:29safer communities. So yeah, that's what
27:32actually I will consider today. Thank
27:33you so much.
27:34>> Thank you, Simma. Next, we have Charles
27:37Briham, senior account manager for the
27:40nonprofit and global organizations team
27:42at ESRI. In his work, he focuses on
27:45sustainable development driving SDGs
27:48efforts. Charles is a geographer with
27:52expertise in the areas of international
27:54development, census and statistics,
27:57geospatial infrastructure, ICT
27:59innovations,
28:01uh image analysis and humanitarian
28:03affairs. Charles has worked for the
28:06United Nations, the World Bank, NASA, uh
28:10the Caribbean community, CARICOM, as
28:12well as in a number of countries in
28:14Central America,
28:17Africa and Asia.
28:19Charles, what are two things you are
28:21most hoping to see from the teams today?
28:29>> Thanks, Taja. Um, I think that the the
28:32two big things are um
28:35were mentioned previously and
28:37articulated by PRI in her in her
28:40overview, but scalability and
28:42practicality, I think, are are two key
28:45things um I'm going to be looking for.
28:47And of course um based on u technology
28:52aspect around AI and uh the aspects of
28:56of data and data models that are feeding
28:59that that model process and I just want
29:02to appreciate all the um the work that
29:05the teams have put into this effort. you
29:08know, someone lived in East Asia before
29:10the road accidents, as Briti mentions,
29:13one every minute. And I think that's
29:15something that we all want to see see
29:17fixed. And my layman's terms, I've
29:20always been an advocate for bumpers
29:24being at the same height globally as one
29:28simple solution. But here today, we're
29:30going to see some really innovative work
29:32and I'm excited for that. Thank you.
29:35>> Thank you, Jos. Next we have Richard
29:38over the CEO and road safety specialist
29:42at Agillysis. He has over two decades of
29:45experience working in the field of road
29:47safety. He specializes in road risk
29:49analysis, enforcement strategy, policy
29:52setting and connected vehicle data. And
29:56uh Richard, the two things um you will
29:59be listening in for in today's speeches.
30:02I think what I've really enjoyed about
30:04this process so far is seeing how people
30:07who are experts in their field but not
30:09necessarily in road safety have you know
30:12taken lots of different approaches to
30:14try and solve a problem that many of us
30:16who've been working in this domain for
30:18for decades um have been trying to fix
30:21using traditional solutions. So I'm
30:23really looking forward to the level of
30:25um innovation and broad thinking and
30:28different strategies um that people have
30:31um delivered. The other thing that I'm
30:33also looking for given that you know we
30:35spend an enormous amount of time
30:36advising governments um across the world
30:39on what strategies look like is how
30:42explainable and understandable the tools
30:45are. We don't want a whole bunch of
30:47different complicated numbers and
30:49algorithms. We need simple results that
30:52um that uh practitioners within
30:54individual countries can use to not only
30:56understand what safe speeds look like
30:58across their networks but also on
31:00individual roads.
31:04>> Thank you Richard.
31:06Next joining us from DC Washington DC
31:09USA Holly Grampek program manager in the
31:12development data group at the World
31:14Bank. Among many things, she is the
31:18founder of the development data
31:19partnership, a partnership between
31:21international organizations and
31:23companies created to facilitate the resp
31:26responsible use of thirdparty data in
31:29international development. She also had
31:32a background in leading the bank's
31:34transport work in East and Southeast
31:37Asia.
31:38Holly, what will be uh you be listening
31:41in for today?
31:44>> Thank you. Actually, my program got its
31:46start using alternative data to evaluate
31:49road safety in the Philippines uh maybe
31:5212 years ago. So, it's really exciting
31:55to see how the field has evolved. Thank
31:58you for the invitation. Uh I will be
32:01looking primarily for utility. Uh do
32:04these applications answer the right
32:06questions that will genuinely lead to a
32:09road safety improvement uh and
32:11reusability. To what extent can these
32:14applications be scaled to other
32:16countries or uh to other organizations
32:18like the World Bank? Thank you.
32:22>> Thank you for the answer. Uh last but
32:26not least, Mika Mukai, principal ID
32:29specialist from the information
32:30department uh information technology
32:33department at the Asian Development
32:35Bank. He manages digital projects or
32:38products supporting sovereign operations
32:40and development projects. He brings
32:43experience across public infrastructure,
32:45IT and management consulting and is
32:48particularly interested in using data,
32:50GIS, earth observation and emerging
32:53technologies to enable cross- sector
32:55solutions and create greater development
32:58impact. Last but not least, what about
33:01you Mikia? What are two things uh that
33:04would make a solution stand out in this
33:07afternoon?
33:08>> Yeah, thank you for introductions. Um
33:11first of all I can congratulations that
33:14we want to get to this point and uh uh
33:17it's a very important matter on our road
33:19safety for our lives and also injury but
33:23also economic impact of um the uh road
33:27is not being able to recon uh to being
33:31able to use form uh for the purposes of
33:34this build. So uh I think it is very um
33:38the great uh theme to uh to challenge on
33:42and uh um uh in this transport is a
33:46geographical information is an mandatory
33:50um aspect of this and then this is a
33:52part of my um uh one of the digital
33:56portfolio. So I'm very interested in see
33:59how everyone is actually approaching
34:01this. So two aspect is from IT
34:04perspective in more scalability
34:07and uh also reusability.
34:10So the scalability mean is that um how
34:14this could be an expand uh in uh regions
34:18and countries but also for um uh another
34:23aspect of a transport or even other
34:25things and the reusability is that um
34:30often come within certain architecture
34:32that um uh the this program or analysis
34:38is an open even like and to be used for
34:41other purposes within fine tune of this.
34:44So uh I think this uh um uh challenge is
34:49pretty variable. So uh I'm looking
34:51forward to seeing this and then uh good
34:53wishes for everyone for your for your
34:55pitch. Thank you.
34:57>> Thank you all for being here today and
35:00lending your keen expert eyes and ears.
35:03Our judges are scoring against five
35:07criteria shown on the screen. Half of
35:10the weight sits on the methodology and
35:13whether the solution can scale to other
35:15countries and the other half is whether
35:17the uh outputs are accurate, actionable
35:21and clear enough for road safety experts
35:24and government planners to actually use.
35:26Just before we start the pitches, let's
35:29again review the rules. Each team has
35:31five minutes to pitch and de demonstrate
35:33followed by five minutes of questions
35:36from the jury. Not for the jury. Um,
35:39when the Q&A start, just lift your Zoom
35:42hand up and I'll moderate the questions.
35:45For the audience, you can pose your
35:46questions to to the Q&A box and the
35:49teams will answer them after their
35:51pitches. And at 5:45, be ready to vote
35:54for your favorite for the audience
35:56choice award. All right, let's begin.
36:02Our first team is Neural City joining us
36:05from India. Brites
36:08and Ash you can take the
36:12>> uh thank you. Uh good afternoon
36:14everyone. Uh we are neural city and this
36:16is safe speed. Uh what we have built is
36:19a model which predicts safe speed for
36:21each road segment. uh speed which is
36:23safe for most uh vulnerable road users
36:26as uh the survival speed uh is different
36:29for different road user. Um hit at 50 a
36:32pedestrian and has around 36% chance of
36:36getting hurt and while uh let's say our
36:39reduction in 10 kilometer uh reduces the
36:43chance to almost 10%. What is different
36:46in our approach is u we read the whole
36:49place and not just the speeds and uh uh
36:53talking about how our overall pipeline
36:57works is we ask three questions mainly
37:01uh who can be hit uh that sets the safe
37:03speed and its gap uh to the posted limit
37:07uh second question is how badly uh so we
37:09use TomTom's real speeds uh through the
37:11injury curves and give the expected harm
37:15And in case when data is thin, a model
37:18predicts the speed that the road
37:19actually invites. And finally, how many
37:21as in the exposure, people nearby and
37:24traffic. And together based on this uh
37:27we create a priority score. Uh and based
37:30on these priority score, we split them
37:33into bands to see how each acts. Since
37:37we don't have any record for crashes but
37:39we have validated through academic
37:42published research we match for for most
37:46of the segments that was we match uh to
37:48close to 2.5% parity and similarly
37:53we do this for we found a road N bridge
37:55in Pune where the police has set the
37:58speed to 30 but and we have predicted
38:01the same as now continuing on with the
38:03demo.
38:07Save demo by neural city. Uh on our
38:11dashboard you can see the road network
38:14for the target target geography. This is
38:16for Maharashtra. The road segments are
38:19color coded. Uh yellow and green being
38:21low priority road segments and orange
38:24and red are high priority. Uh color
38:28coding is based on the priority score as
38:30described earlier. Uh the dashboard also
38:33gives you a summary of road segments
38:35which have been analyzed, total length
38:37of road network, number of roads which
38:39are in high priority uh uh zone. And the
38:42dashboard for Thailand is also very
38:44similar. And if someone wants to look at
38:47the priority map only, they can like
38:50priority segment, they can go to our
38:52priority map. Uh
38:55the urgent road segments are uh
38:58shortlisted. You can zoom in and click
39:01uh on the road segment which requires
39:03action. It will give you the road name,
39:06what is the priority score, current
39:08speed limit versus what we are
39:10suggesting, primary reason. So here for
39:12this segment, a potential pedestrian
39:14impact uh a road seger
39:18is also available and it also gives you
39:21uh based on the imagery what features
39:23have been identified. If you look at the
39:26imagery, so this is for the same road
39:28segment. Uh we can see two wheelers
39:32present, pedestrians are also there and
39:35other commercial vehicles are also
39:38present. If one wants to look at other
39:40road segments, they can just move around
39:42and click on those. It will again give
39:44you a similar summary. So one
39:47intervention that we suggest obviously
39:49is lowering speed. Other uh for other
39:52engineering suggestions we can go to
39:54what to do tab where if you select a
39:58road a segment it will give me an
40:00engineering treatment let's say for this
40:02road it is suggesting a raised
40:04pedestrian crossing why because lot of
40:06people live nearby and if I have to look
40:09at the density I can just scroll down
40:12people living within 300 m is 51,000
40:15it's uh urban land use around this road
40:18and it also shows me uh what is the
40:21evidence we are using and percentage
40:24reduction in fatal and serious injury
40:26the speed is low. So this is how our
40:30dashboard lets anyone move around uh
40:33very uh intuitively around the map uh
40:36select target geographies or areas and
40:39see where to act and what to do.
40:45>> Uh okay thank you. So continuing forward
40:48uh in cases where there are probe speeds
40:50available uh we use the probe speeds at
40:53the target variable and in cases where
40:55there is no probe speed available they
40:57are also just using photo vision model
41:00extracts the IR road attributes and we
41:03can still suggest the a safe speed
41:07>> and thank you that was just on time
41:11thank you thank you Ash very good
41:14presentation now Do we have questions uh
41:18from the jury?
41:20>> Yeah, I was going to ask um if they
41:22could explain the the vision model a bit
41:24more.
41:25>> Sure. Uh so uh there is already been a
41:29study uh by ThyRaph based on Thailand
41:32where what they did is let's say for
41:34around 2,000 they have 2,000 images and
41:37they therap authority based let's say 52
41:41attributes are there and they first
41:44manually annotate them and then uh they
41:47kind of train simple models both VLM and
41:50YOLO based models as well and what we
41:53did was is basically built on top of
41:55that apart uh So uh we used let's say
42:00open sour we used let's say gemini for
42:02annotation and for we extracted the
42:05mapillary images and then kind of rated
42:10along those 42 IRA attributes. Uh that
42:14was the core of that vision model.
42:19>> Great. I really like the visualizations
42:22and how accessible you make the
42:23different data layers. I was wondering
42:25if you could explain a little more how
42:28the treatments are determined, which
42:30data go into that and where the
42:32treatments come from.
42:34>> Sure. Sure. U actually can you move into
42:36the method part? Yeah, sure. So, uh our
42:40uh target variable is mostly probe
42:42speeds because that is where we know
42:44what's the actual uh let's say flow of
42:47traffic is. So that is the let's you can
42:50say target variable and for each for
42:52each road segments uh we have let's as
42:54you can see around five models starting
42:57with the main model is basically let's
43:00say a tree based or gradient boosted
43:01models because u that is the main model
43:05and uh for other information let's say
43:07for um places like whether it's a
43:10roundabout or it's a junction all those
43:13things so different uh let's say
43:14attributes are calculated from different
43:17models and uh uh for and from street
43:22images let's say we extract the IRF
43:25based attributes and so on and as you
43:28can see uh once all of these are done uh
43:31then we kind of uh uh create this
43:35priority score based on uh that
43:38>> yeah and just to answer if by treatment
43:41you mean the engineering treatment side
43:42how we are suggesting and other things
43:45>> there we are using the AL our research
43:48from 2019 where they have done uh
43:51researches uh and was what type of
43:55construction changes calming measures
43:57can be introduced and second study was a
43:59world bank study from 2017 where they do
44:03the I think again in Thailand they did
44:05the uh road safety research regarding uh
44:08uh infrastructure changes based on the
44:11conclusions in those researches whatever
44:14intersections we find based on our data
44:16for indeed individual segment we suggest
44:18uh intervention.
44:21>> Um I've question that's okay.
44:24>> That's okay. We have one more minute.
44:27>> Okay. Um I noticed that um quite often
44:30you were predicting 30 km an hour um
44:32wherever you're in highly populated
44:35areas but many of those roads may have
44:37the level of infrastructure required
44:39where they've um separated road users.
44:41I'm interested in you know what you'd
44:43want to do to further improve the model
44:45to actually identify roads where the
44:48speed limit could be higher.
44:51>> Yeah. So currently uh Richard as you s
44:54highlighted we are focusing on 30 based
44:57on our own experience let's say in India
44:59even it's a national highway and
45:00expressway you can find two wheelers
45:03flying in any lane right so that is why
45:05we have overfitted today or or it's it's
45:08it's a bias towards safety of a
45:11pedestrian right that is obviously a
45:13bias today what we have to do is
45:16identify if people are in correct lanes
45:18there are barriers uh uh there if se
45:21segregated ing people or if it's a no
45:23two wheeler zone two wheelers are not
45:25there these are some things intuitively
45:27which I can think of pratish model wise
45:30if anything you want to add
45:31>> uh yeah true so we have a variable let's
45:34say traffic mixup which kind of says
45:36what kind of uh primary road users are
45:39if let's say we spot a two-heer and
45:41let's say pedestrian then we
45:43automatically reduce the score so uh
45:45same thing which aa said
45:48>> thank you so much we're at time but If
45:51you would like to continue the
45:53conversation or your answer on the chat
45:55side, that's completely allowed. So, and
45:58poor people, you can ask questions in in
46:01the chat. Now, um moving on to our next
46:04team. Our second team is Beachroot
46:09joining us from the Washington DC USA.
46:12And while Hanoi gets ready, Beachroot is
46:16a review cue for speed limits. It ranks
46:20the roads where the posted limit may be
46:22wrong and for each one names the
46:24instrument that sets the limit and the
46:26office that can change it. Hanoi, the
46:30floor is yours once you're ready.
46:34>> Okay. Can you see my screen and hear me?
46:37>> Yes, you are perfect.
46:39>> Okay, so I will start. Um, hi, thank you
46:42for having me. Um, I'm Hanuan and this
46:44is Speed Choose a review queue for speed
46:46limits. It puts traffic speed, poster
46:48limits, and roadside exposure together
46:51and runs the roads an agency should look
46:53like first. So, a ranking does not tell
46:56an agency what to do because it doesn't
46:58say which rules applies or who can
47:00change the number. On Gena Pac, the data
47:03set marks 180 kilometers at 120. The
47:08general selling is 100. So, every one of
47:11those kilometers needs a notice naming
47:13it. The notice I found names eight. From
47:16the rest, I could not find one. So,
47:19which rows should an agency review
47:21first? Every road gets a reference
47:24speed. The speed a person could survive
47:26being hit at road contest sets it. Maps,
47:29schools, and markets tell me who is
47:31exposed and where a street photo shows
47:34the road. Clearly, a v a vision model
47:36upstairs. The score combines three
47:39things. How far traffic runs above the
47:41reference. How far the posted limits
47:44sits above it and the exposure weighted
47:4645, 35 and 20. Keeping the first two
47:50part matters because a road can be
47:52driven entirely within its limits and
47:55still carry a limit nobody could
47:57survive. The weights are my own choice.
47:59Uh here is one road in map I submitted.
48:02Speed choose scores 14,700 road segments
48:05across Thailand and Mahab Russia within
48:08a supply network of about 70,000
48:10segments. Here I filter the Thailand and
48:12Lita corridors so neighboring segments
48:14can be reviewed together. I open the
48:16force rack corridor. All five flag
48:18segments are critical. Traffic speed
48:20exceed their reference speed by at least
48:2330 km per hour. The suggested action
48:25started with lowering the limit. The
48:27lift image is 22.6 km from the
48:30inspection point. This notice list an
48:33exception of the Bangkok section of this
48:35road, not the whole road. The selected
48:37corridor lies mainly farther north. The
48:40documents geographic scope therefore
48:42needs checking before applying it. Here
48:44the legal register contains 90 entries.
48:4617 are available. Two are explicitly
48:48marked missing. It keeps the source and
48:50gaps visible. The same pipeline also
48:53runs on Maharashtra. Here the top ranked
48:55corridor has no imagery. So the
48:57assessment uses map road conditions.
48:59Confidence is shown separately from
49:01priority. The sity responsible for
49:03changing this limit remains unverified.
49:05Here's the map submitted in ADB's RTS
49:07environment. Selecting a corridor shows
49:10its rank, suggested first action and
49:12indicative course. Here reviewers select
49:14corridors for a few visits with a
49:16checklist of what to report on site. The
49:19selected location download as a CSV
49:21directly from the browser without a
49:23separate account or desktop. GS imagery
49:26changed severity band of 1736 segments.
49:311552 moved on. It also moves 39 segments
49:34from the bottom two priority tiers into
49:37priority one. 10 are modeling class
49:39where image to road matching needs
49:41rechecking before action. Imag covers
49:44about a quarter of square segments or 5%
49:47of the supply network. Around 4,000
49:49square segments lie within 150 m of a
49:53school or market.
49:55So how far can you trust IQ? I rated 24
49:58photos blind before seeing which ones
50:00the model had flagged and we agree on 19
50:03my own rating. So it measures
50:05consistency rather than accuracy. A
50:08photo stands for about 3 kilometers of
50:10road. I also checked it against a field
50:13the model never sees. On the weights I
50:15ran 2,000 combinations and 800 and 481
50:20of the 1473 priority one segments
50:24survived almost all of them. Nether
50:26outcome check held up. One vanished once
50:29I controlled for network size. The other
50:32covered only like 10 provices. So
50:35segment level crash validation is still
50:37missing. Um the next step is a pilot
50:40with a transport department to field
50:43check a sample of the flag corridors
50:45that gives a measure of false positive
50:48rates I cannot produce on my own and it
50:50shows where the next day imagery should
50:52go. The framework itself travels two of
50:55three factors need only open map data
50:57but the but the legal layer had to be
51:00rebuked locally. Um when command check
51:0344 of my 61 published the figures and
51:06names the 17 it cannot
51:09so sorry
51:13wait sorry uh I lost to one page of my
51:17slides but it's okay uh I just want to
51:19mention so our first meeting gave speed
51:21choose is it tuktac logo you will see it
51:24uh so later I went looking for a
51:25survivable speed for the passenger
51:27inside one and the literature I reviewed
51:30doesn't have that number So I keep gap
51:32visible and this is what uh speed choose
51:35for giving a user somewhere to start and
51:38saying what still did checking. So
51:40here's the end of my presentation. Thank
51:42you.
51:43>> My goodness honey just on time. Well
51:46done. All right let's go for the for the
51:51judges questions. Um,
51:54Thema, can I put you on the spot since
51:57um,
51:59Perfect? Fimma, go first.
52:01>> Yeah. [laughter]
52:03Yes. Uh, very interesting. Uh, my
52:07question to you is like, uh, how do you
52:09think a planner can use your map on
52:12their own? Um, because I see like, yeah,
52:15do you think like it is easy to make it,
52:18um, for a planner?
52:21Yeah, I think um I think a planner can
52:24open the queue in a browser and choose a
52:26corridor for review because it's based
52:29on uh it's based on my GitHub page. It's
52:32not u like you need a professional RG
52:35skills or something else. So the Q shows
52:38the rank and the suggested first action
52:42like each card gives an an inspection
52:44point and cost here for planning the
52:47visit. So um I also show the available
52:50legal information and the name the
52:52office that can change the limit. Yeah.
52:55So so yeah so the planner can download
52:58the selected location as a CSV without
53:01the uh GS account or something else.
53:05>> Okay.
53:07>> Thank you.
53:09>> Next to uh to remember I don't see your
53:12hand.
53:14>> Um can I ask a quick questions? Um
53:17>> absolutely ma go ahead.
53:18>> Yeah. Hi. Is this I I trust in this is
53:22based on like kind of the data you
53:24acquired. it is this like uh the data
53:27need to be refreshed because it's just
53:30look to me and a little bit more like
53:31kind of static the the data sets and
53:35then I know models but um I was
53:38wondering how it could be more like uh
53:41operational
53:42um to to um to to use this and
53:45operationally for a long period of time.
53:49>> Well uh thank you for your question. So
53:52uh I have not set an update schedule for
53:55it. So the 20s and like maybe some
53:58agency owner to become like operational.
54:03>> Okay,
54:04>> that's it.
54:05>> Okay. Um does it require does it does it
54:09require to refresh the data or is this a
54:12more one kind of one time?
54:15Well, I think um
54:18I think yeah, I think it should be like
54:22um based on like it needs to refresh
54:25data like to remain useful like rather
54:28than like just being a one-time
54:30assessment for it like um because I
54:33think like new traffic speed data would
54:36let the agency updates it it's ranking
54:39for it and uh updated road maps could
54:42capture changes in the network
54:44surrounding ings and I see there is a
54:47question in the chat box asked so what
54:49is the innovation question so that is my
54:52answer I want to say say it here so uh I
54:55think the the innovation question is
54:57like the unsafe speeds uh based on uh on
55:01the road based on like the the mixed uh
55:05the mixed condition by by cars by like
55:09by years or changings time changings and
55:12different condition of people and cars.
55:16So, I think that's why we need to update
55:18it. It's like for ranking or something
55:21else.
55:22>> Okay. Thank you.
55:23>> Thank you so much.
55:25>> We still have time for Holl's questions.
55:281 minute 15 seconds. Thank you.
55:30>> Great. Thank you. I I love that you you
55:34have a GitHub and that it runs from the
55:35browser. That's great for scalability.
55:38Uh, I I wasn't sure if I caught it
55:40correctly in your presentation, but did
55:42you have your results verified by
55:45persons in Thailand? Is that what you
55:47said?
55:49>> Uh, yeah. Uh, I I didn't have my results
55:53independently verified by people in
55:55Thailand, but I just talked to my friend
55:58from Thailand to teach me how to read
56:02the the the name in Thai.
56:05>> Okay. Thank you. Yeah. So that is
56:08remains like a gap here.
56:15That's for further development. All
56:17right. Craig, is there any new
56:22questions? Otherwise, we
56:26we can uh say thank you to Hanui and you
56:29can still keep on answering questions on
56:32the chat side. Thank you so much. Now,
56:35let's move on to our third team. And our
56:39third team today is Gaudi uh joining us
56:43from Singapore. And while Andy gets
56:46ready, their solution's main idea is to
56:48apply AI at the data preparation stage
56:51to feed into a strong IRA model that
56:55outputs safety ratings for road users at
56:58scale.
57:00Go ahead, Andy. The the floor is yours.
57:04Hi everyone, good afternoon and thank
57:06you for having us and a big thank you to
57:09ADB and partners for organizing this
57:12program. Our topic is about getting
57:15crowdsource read imagery or road level
57:18imagery into IRA coding and generating
57:23star rating street safety rating using
57:27AI. It's a long title so let me unpack.
57:31Here is our team member from team Gaudi.
57:34We are young team of data scientists and
57:36AI application engineer. Most
57:38importantly, we have direct access to
57:40engineering domain experts within the
57:43group company of ID and Namcoi Japan.
57:48We have built a AI data processing
57:50pipeline that prepares data from three
57:53sources,
57:54creates input data compatible to road
57:58rating model, IRA, and outputs
58:01internationally recognized safe road
58:03rating per road segment. Uh, IRA or
58:07International Road Assessment Program is
58:09a UK registered charity established to
58:12help tackle the the devastating social
58:15and economic cost of road crashes. With
58:18the data, we also created three
58:20interactive RGis online dashboard for
58:23road safety planners to use and safety
58:26city authorities.
58:30Existing IRA input data coding stage is
58:33manual and labor intensive.
58:36Um it's common in engineering survey and
58:39is no surprises here. So for every 100
58:42mters of road segment data road
58:45attributes are cor collected and input
58:47into spreadsheet maybe with some
58:50software help. Finally, the most
58:52important stage is have a team of people
58:54in the picture here is 919 to perform
58:57peer-to-peer coding review.
59:02Our main idea is to apply AI
59:06at this data preparation stage to create
59:09high quality data input that fit into
59:12the strong IRA model and generate star
59:15rating for road segments.
59:19Uh this slide shows our high level block
59:22diagram. We hope the audience appreciate
59:25our team's effort. It looks easy on the
59:27slide. Uh but we have um spent a lot of
59:30challenges and method putting this into
59:32a pipeline.
59:36Our method achieve cost effectiveness
59:39for member countries. Um so it's easy to
59:42set up
59:43simplified uh unified platform
59:46lightweight processing very important.
59:48Uh it takes about 40 minutes with GPU
59:50just 3 hours on a regular laptop to
59:53process 100 and uh sorry thousand plus
59:57street view imagery and most importantly
1:00:00the output is
1:00:03open data.
1:00:06So these are the sample imagery. Um and
1:00:09I would like to give us show a scenario
1:00:14where the road planner
1:00:18using our dashboard can actually easily
1:00:21see the coded region. Uh red color is
1:00:25poor to road users. There are four types
1:00:27of road user. uh green is good so they
1:00:30can quickly locate the problem uh and
1:00:33then check the corresponding street view
1:00:35imagery. The key road segment
1:00:38infrastructure attributes are on the
1:00:40left hand side.
1:00:42Once they and it's
1:00:48once they identify the road segment of
1:00:50interest, they can then gather and input
1:00:53uh put it select and gather them into
1:00:55the simulator which they can then tweak
1:00:58the attributes for examples road speed
1:01:02uh different road attributes to get
1:01:04better
1:01:06outputs.
1:01:07Better star rating. So better star
1:01:09rating are targeted to road users like
1:01:11pedestrian vehicle occupants,
1:01:15motorcyclist and cyclists. So let me go
1:01:17back to the slide.
1:01:21Right. So from the from from the model
1:01:25we are aware that um the misconception
1:01:27is
1:01:29road speed is one of the element to
1:01:31tweak. In fact, the road planner can do
1:01:33can maintain road capacity by keeping
1:01:35the road speed unchanged but changing
1:01:38improving the road infrastructure.
1:01:41Scalability rapidity yes with these four
1:01:44ingredients.
1:01:46Other benefits are AI for good dual use
1:01:48road assessment program
1:01:51can use the data input
1:01:53for close. Our team would like to
1:01:55continue develop this prototype um by
1:01:58fixing some of the problems with the
1:02:00dashboard and ultimately we would like
1:02:03to pipeline this um pilot this AI
1:02:06pipeline with um ADB
1:02:09at Thailand. Thank you.
1:02:12>> Thank you so much Andy and um let's go
1:02:16to our jury Richard.
1:02:19>> Uh thank you very much for the uh the
1:02:21presentation there. um you know the the
1:02:24IRA community um covers many researchers
1:02:28working across the globe to try and um
1:02:30fix the coding issue that you talked
1:02:32about using AI. Um I definitely
1:02:35recommend that you investigate the AI
1:02:37rap program. Um and I'm sure they'd love
1:02:39to speak to you about that. But
1:02:41specifically about this, you know, doing
1:02:43that work is fantastic. How does that
1:02:45translate itself into a safe speed
1:02:48score? Um you've got the star ratings
1:02:51but how do we understand what the safe
1:02:53speed limit should be
1:02:56>> right um in in fact we think of this way
1:02:59um following the IRA methodology um
1:03:05when we detected the road segments um
1:03:09mark for example to to different type
1:03:12user for example we are targeting
1:03:14vehicle occupants for that particular
1:03:17road segment Um we would it's with the
1:03:21road attributes we can get the existing
1:03:24uh rating. So the road planner we can
1:03:26tweak the road attributes to achieve
1:03:30better score. Uh what we found is that
1:03:33tweaking the operating speed uh actually
1:03:36doesn't contribute much. So actually
1:03:39it's a to us it's a misconception.
1:03:42uh in fact adding let's say channelized
1:03:44pedestrian crossing are in in fact make
1:03:48it safer for the pedestrian while
1:03:50keeping the vehicle speed unchanged.
1:03:55So so um
1:03:58thank you
1:04:01>> thank you Sema. Yeah.
1:04:05>> Yeah. I really like the technicality,
1:04:08uh, Andy, but [clears throat] uh, I
1:04:09still have a question on the processing
1:04:11because what I noticed, you mentioned
1:04:13that processing is slow about 40 minutes
1:04:16per thousand images on a GPU, right? So
1:04:19uh what would it take to in time and
1:04:23cost uh to run this across a full
1:04:26national network and how accurate is the
1:04:29automated coding where you used dummy
1:04:31values?
1:04:32>> Oh yes uh yes uh yeah so in fact it's
1:04:35not slow it's actually very fast um for
1:04:38for this uh 40 minutes for thousand plus
1:04:40images are very fast. Um in fact we
1:04:44extropolate to do a 100 meter segment
1:04:48uh it just uh take 3 hours for example
1:04:51uh that is uh uh in fact fast. So on the
1:04:55accuracy uh that is where our team uh
1:04:58would need to be certified uh and
1:05:00actually uh actually I will show the
1:05:02limitation. So there are limitations in
1:05:04the current method. Uh for example uh
1:05:07the street view imagery uh we do not
1:05:10capture enough um is mostly uh ahead
1:05:15looking so we need to look at left and
1:05:16right for the road attributes. So this
1:05:19need to be improved. Um the other aspect
1:05:23would be the um to do a once the road
1:05:28segment of interest is identified we
1:05:29need to actually not rely on mapillary
1:05:32but actually do our own drive to capture
1:05:35the road imagery um and explore maybe LA
1:05:41uh the coding is uh we need to get
1:05:43certified uh is it's uh
1:05:47it's quite hard that part I would Okay.
1:05:51>> Yeah.
1:05:52Okay. Thank you.
1:05:54>> Thank you so much. And uh Charles.
1:06:00>> Yeah. Uh thanks for sharing this work.
1:06:03Um, you know, my questions were
1:06:05partially answered by um um by you from
1:06:10from Richard and the previous uh our
1:06:12previous drawer, but just maybe just add
1:06:15something. If you were going to provide
1:06:17a um uh pitch to policy makers in the
1:06:21last 30 seconds, what would that be?
1:06:27>> Right. For policy maker, um we should do
1:06:29a rapid um in engineering require rapid
1:06:32evaluation. So with our system we can
1:06:34give it in fact free of charge a low
1:06:37cost countrywide Thailand or Madrasa uh
1:06:41and give them marking uh show them show
1:06:44them the road segments
1:06:47of risk to the four types of user and
1:06:50start that conversation
1:06:52and then use the recognize uh introduce
1:06:55the recognized IRA methodology to guide
1:06:57them where to spend the investment in
1:07:00the infrastructure road infrastructure.
1:07:02and then do the simulation to show that
1:07:04with that how much life you can save. So
1:07:08it's a conversational starter with the
1:07:10interactive uh easy to use dashboard.
1:07:15>> Thank you so much team Cody and thank
1:07:18you to the jury for the good questions
1:07:20and let's move on to our next team. Our
1:07:24fourth and final team today is 5S
1:07:28joining us from Georgia and India. While
1:07:31Kentaran Viva are getting ready, a
1:07:33little bit about their approach. 5S sets
1:07:3630 kilometers per hour. The speed with
1:07:3990% survival rate for vulnerable road
1:07:42users as the baseline safe speed and
1:07:45raises it only when physical separation
1:07:48is confirmed via AI or other means
1:07:51offering a safer and explainable
1:07:53framework that naturally scales even in
1:07:56long term as innovation continues and
1:07:58data accumulates.
1:08:01Can Tara and Vivc, are you ready?
1:08:06>> Yes.
1:08:07>> Can you see my screen?
1:08:10>> Yes.
1:08:12>> Perfect. Your time starts now.
1:08:15>> Okay.
1:08:17Thank you for the introduction. We are
1:08:19team 5s.
1:08:21A safe speed can be the highest speed at
1:08:23which a human body can survive 90% of
1:08:26the time.
1:08:28In the safe system approach, it varies
1:08:30by potential collision type. A car
1:08:32hitting a person 30 km/h. A side impact
1:08:36at a junction 50. Head on 70. A medium
1:08:40with no pedestrians 80 to 100. And
1:08:43motorcyclist in the lane with cars 30.
1:08:46Again,
1:08:48now imagine, does pedestrian absence in
1:08:50street photos mean nobody walks there?
1:08:55The answer is no. But the opposite is
1:08:57true. That's why PIS treats as if
1:09:01pulable road users are there unless
1:09:04evidence suggests otherwise.
1:09:07We have three design principles.
1:09:10Measure what is measurable and make
1:09:12measurable what is not. So avoid
1:09:14compounding big assumptions. And unless
1:09:17evidence justifies more, the safe speed
1:09:19remains 30 km/h.
1:09:22Our speed safety score considers the gap
1:09:25between the current process speed and
1:09:27the safe speed. People on foot and
1:09:29confidence in the data.
1:09:33For example, when the speed limit is 60
1:09:35and the safe speed is 30, the speed gap
1:09:38is 1/2.
1:09:40People on foot and reliable data score
1:09:42one each at maximum and higher number
1:09:45indicates higher priority.
1:09:495S uses nine independent map layers.
1:09:52Four of them caps safe speed and the
1:09:55other set re the action priority.
1:09:58Now let's go through each layer.
1:10:01Layer one, today's poster speed limits.
1:10:05Layer two, road classification. For
1:10:08example, can pedestrians use each road?
1:10:12Layer three, the default safe speed of
1:10:1530 km/h.
1:10:17In our AI layer, PBS raises safe speed
1:10:20when vulnerable road users are protected
1:10:22by physical separations.
1:10:24When map photos indicates vulnerable
1:10:26road user existence, 30 becomes the cap.
1:10:32Likewise, the safe speed is 50 at
1:10:34maximum at junctions.
1:10:37To protect children, 5S traces where
1:10:39they walk in 3 minutes from school and
1:10:42caps those at 30.
1:10:44From layer 79, it calculates how far the
1:10:47existing uh speed limit is above the
1:10:49safe speed, determines potential
1:10:51pedestrian abundance based on population
1:10:53density, calculate safe speed score, and
1:10:56rank segments for practitioners.
1:10:59Here are two demos.
1:11:03>> Which roads need action first and why?
1:11:05[music] The 5S dashboard answers both by
1:11:07mapping road class, daily average
1:11:09traffic volume, [music] speed limit,
1:11:11safe speed, and each segment's priority.
1:11:14Click a top priority road. The higher
1:11:16the safety score, the sooner it needs
1:11:18action. [music] Every segment has a
1:11:20score explanation. Here, the current
1:11:22speed limit is 50 kmh above the safe
1:11:25speed, [music] and the data is reliable.
1:11:27On this low priority road, the gap, on
1:11:30the other hand, is only 15 kmh, [music]
1:11:33and the data is less cert. So before a
1:11:35site visit or a meeting with the traffic
1:11:37police, you know which road to act on
1:11:39first [music] and exactly why.
1:11:42Our second tool is the safer speeds tool
1:11:45built with ArcGIS experience builder.
1:11:48You can estimate the benefit [music]
1:11:49cost ratio of safety treatments per road
1:11:52segment or for the whole network using
1:11:54three tabs explorer counter measures and
1:11:56results. According to [music] IRA, a
1:11:59median barrier reduces head-on
1:12:01collisions by 90%. So as an example,
1:12:04let's see what would happen if we
1:12:05install median barriers to [music] the
1:12:07top priority roads which are the top 3%.
1:12:10The analysis period, cost per kilometer,
1:12:13crash reduction factor and other inputs
1:12:15come from IRAP documentation [music] or
1:12:17can be configured for local conditions.
1:12:20Results show the treatment cost, deaths,
1:12:22and serious injuries before and after,
1:12:25and the benefit cost ratio, with which
1:12:27decision makers [music] see in minutes,
1:12:28where a treatment pays off, and direct
1:12:30budgets to save the most lives. And what
1:12:33about individual segments? No worries.
1:12:35[music]
1:12:35The explorer tab shows the same figures
1:12:38for each segment. More counter measures
1:12:40can be added in the pilot
1:12:41implementation.
1:12:46Sometimes mapillary detects physical
1:12:48barriers but not whether they stand
1:12:51between the sidewalk and cars and we
1:12:54have ideas to fix this with a pilot. We
1:12:57like to add a visual language model
1:12:59layer to tune it on the field visit in
1:13:02India and Thailand.
1:13:04BS it's a true safe system by design
1:13:07where missing data never raises safe
1:13:09speed. every result is expandable and uh
1:13:13it can be tuned for local policies.
1:13:16Thank you.
1:13:17>> Thank you so much Gentara for the
1:13:20excellent pitch.
1:13:23Now let's move on to the jury.
1:13:27Can I call um Richard first?
1:13:35>> Uh yeah, that was a really good uh
1:13:37presentation. I I love the way that you
1:13:39um uh created that decisionmaking tool
1:13:42at the end there. Um uh how reliable um
1:13:48do you think that's going to be for road
1:13:50authorities? You know, you've talked
1:13:51about how you'd want to make
1:13:53improvements. You talked about some of
1:13:55the um computer vision challenges that
1:13:57you had in there. What's the one thing
1:13:59that you think you'd want to do to make
1:14:01sure that that decision-m tool is as
1:14:04robust as possible?
1:14:07That's a good question.
1:14:09I think our system is a relatively
1:14:12reliable when it comes to
1:14:14uh false negative rate but uh maybe now
1:14:21it's it's sometimes it can be like a too
1:14:25conservative.
1:14:27So I think uh the point is uh the
1:14:31reliability will in only probably
1:14:34increase when uh as innovation new
1:14:38innovation comes in like visual language
1:14:40model because uh our strength I believe
1:14:43is a design
1:14:45uh that we start from uh 30 kilometers
1:14:48per hour but when it come to reliability
1:14:50we need to test it still. Yeah, no,
1:14:53that's an entirely reasonable approach.
1:14:54Start from 13 work up and certainly
1:14:56that's what a lot of the other teams did
1:14:57as well. Thank you.
1:15:01>> Thank you.
1:15:03Mikio, would you like to go next?
1:15:06>> Yes. Um yeah, thank you for uh the
1:15:09presentation. I think it'll put a pretty
1:15:11story line uh well together from um from
1:15:15the the accent type to like end to uh on
1:15:19the demo on the counter measures. So my
1:15:22my question is that is this a single
1:15:24model serving for different counter
1:15:27measures like uh the sidewalks barrier
1:15:31and a different way of uh preventing
1:15:34accidents and that's the question number
1:15:37one and then also question number two is
1:15:40for the uh calculating the benefit
1:15:45um because
1:15:47the different uh priority roles and has
1:15:51a different traffic numbers and so what
1:15:54does uh really uh actually uh or your
1:15:59approach to actually compare those the
1:16:02uh absolute number to the more um
1:16:06comparative numbers. Thank you.
1:16:09>> Okay, so for the first questions we
1:16:12could add more counter measures like
1:16:15rumble strips or other stuff. We just
1:16:17didn't have time to add everything but
1:16:21uh we need that's certainly we need to
1:16:23look at uh something we need to look at
1:16:25uh in the implementation phase
1:16:27>> and uh second question sorry why was the
1:16:30question
1:16:30>> sorry the f sorry first one so is that
1:16:33you actually using a single model or do
1:16:36you have to have a different model for
1:16:37the different measures
1:16:40>> we we need to have different model but
1:16:42in this example uh we use the crash
1:16:44modification factor from IRA
1:16:47documentation. So if each country has a
1:16:50better uh crash modification factor,
1:16:52crash reduction factor, they can use it
1:16:55>> and also the cost estimate uh is as it
1:16:59can be as accurate as the input. So if
1:17:01they I think they need to have some a
1:17:05need to calibrate a to be more accurate.
1:17:08>> Okay.
1:17:10And my second question was like how you
1:17:12compare this benefit uh in a more
1:17:16priority road and then a more less
1:17:18priority road and where you have a
1:17:20absolute traffic number is different.
1:17:23How do you how do you actually compare
1:17:25this? Yeah,
1:17:26>> thank you.
1:17:26>> Uh we we can compare across uh the
1:17:30priorities. So in the example we only
1:17:33show top priorities but actually there's
1:17:35a menu called every all streets. So as
1:17:38long as they have a absolute number
1:17:41relative number of uh a detail list that
1:17:44they can compare across priorities.
1:17:46>> Okay. And the but uh when you compare
1:17:50the benefit the uh absolute number is
1:17:54higher in anyway in a priority street
1:17:56because more traffic compared to uh the
1:18:00the less priority streets. So
1:18:03>> um
1:18:05>> that is not necessarily true because in
1:18:06our model our model was based on score
1:18:09based on safe system approach. So we
1:18:11didn't consider uh traffic volume
1:18:15>> as a input score because that's a safe
1:18:18system uh system is not a retrospective.
1:18:22>> Okay. Thank you.
1:18:24>> All right. I think we're at time now.
1:18:27The Q&A has ended, but um we can
1:18:30continue continue the conversation and
1:18:33you can continue answering the questions
1:18:35on the Q&A side. Thank you. Thank you
1:18:40all. Thank you 5S. Um and thank you to
1:18:43the jury for the great question. That
1:18:45brings us to the end of all of our
1:18:48pictures today. And uh that was
1:18:51remarkable set of solutions. It's not
1:18:54easy to pitch during a live show. Um,
1:18:57our jury will now withdraw to deliberate
1:19:00in a separate room and will return
1:19:03later. Uh, they will surely have a
1:19:06difficult decision to make while the
1:19:09jury deliberates. We're turning to the
1:19:12audience. This is your moment. The
1:19:15audience choice Paul uh should be open
1:19:19any moment now. Um, which solution do
1:19:23you think would have the greatest
1:19:25potential for impact? After you have
1:19:28voted, you can join the conversation on
1:19:30the chat side.
1:19:34And while we are moving the jury, um,
1:19:40now let me introduce our next section.
1:19:44Every innovation eventually meets the
1:19:46same question. What what does it take to
1:19:49move from promising pilot to something
1:19:51governments actually use? To explore
1:19:54that, I'm handing over to Johanna Borke,
1:19:57the CEO of Greater Than, an
1:19:59international organization that delivers
1:20:01predictive AI that enables organizations
1:20:04to understand and act on crash
1:20:07probability and climate impact. Johanna,
1:20:11thank you for moderating today. Um, over
1:20:15to you. Thank you very much. You can all
1:20:18hear me I hope
1:20:20>> perfectly. Thank you Johanna.
1:20:22>> Super. So good morning, good afternoon,
1:20:25good evening. As Tasha said, I am
1:20:27Johanna Fiki and I'm the CEO at Greater
1:20:30Than and I am absolutely delighted to be
1:20:33moderating today's panel session uh as
1:20:35part of this AI for safer roads
1:20:38innovation challenge. Uh, as Tasha said,
1:20:40at Greater Than, we specialize in
1:20:42AIdriven risk intelligence for road
1:20:44safety and sustainability where our AI
1:20:47uses GPS data to predict crashes and and
1:20:50climate impact. So, when I was asked to
1:20:53moderate today's session on real world
1:20:55use cases on how technology and AI are
1:20:58driving tangible impact in the transport
1:21:01sector today, I immediately said yes.
1:21:03It's something that's very close to my
1:21:05heart and just a perfect match. So
1:21:08before we we kick off, I want to sort of
1:21:10say that when we talk about road safety,
1:21:12we're actually addressing one of the
1:21:14most urgent global challenges. Uh but in
1:21:16today's panel, we're going to be framing
1:21:18that challenge into an opportunity
1:21:20instead because we're going to be
1:21:21looking at how AI is delivering real and
1:21:25measurable safety outcomes on the ground
1:21:28right now. We're actually moving past
1:21:30the theoretical potential of it to
1:21:33showcase some concrete use cases across
1:21:36mobility.
1:21:38So to walk us through some of these use
1:21:40cases, I'm thrilled to be joined by
1:21:42three distinguished experts. Uh Robert
1:21:44Marty, a data scientist with the
1:21:47development impact department at the
1:21:49World Bank. Rob's work at the IE connect
1:21:52for impact program focuses on rigorous
1:21:55impact evaluations of transport
1:21:58investments leveraging telematics data,
1:22:00CCTV and crowdsource data amongst others
1:22:04to inform safety efforts. Pingim uh is a
1:22:08mobility strategist and associate at
1:22:10Crossroads New Zealand and a principal
1:22:12advisor uh for road safety at Oakuckland
1:22:15Council. Ping brings over two decades of
1:22:18experience in safe and sustainable
1:22:20mobility. She led Auckland's
1:22:22award-winning vision zero strategy and
1:22:25today specializes in using data and
1:22:28predictive modeling for advanced speed
1:22:30management. And Liz Waller who's head of
1:22:34road safety as at Trans Urban. Liz uh
1:22:38leads safety across one of the world's
1:22:40largest toll road operators in
1:22:42Australia, the US and Canada. and her
1:22:44work spans across embedding safe systems
1:22:47approach in designing and building new
1:22:49roads, operating and maintaining assets
1:22:52and developing and communicating driver
1:22:55behavior change programs. So given that
1:22:58we only have 30 minutes with our
1:22:59experts, our session is going to be
1:23:01packed. So first each panelist will
1:23:04share a brief 3inut snapshot of their
1:23:06initiative. Following their
1:23:08presentation, we will have specific
1:23:10follow-up discussion around these use
1:23:12cases and then we will take questions
1:23:14from you, our audience, before we wrap
1:23:16up. So, please feel free to commit um to
1:23:19submit your questions in the Q&A box at
1:23:22any time during the presentations. And
1:23:24with that said, without further ado,
1:23:26let's dive right into our use cases. So,
1:23:28Rob, the floor is yours.
1:23:32>> Perfect. Thanks so much. Let me share my
1:23:34screen.
1:23:37uh and put it in presentation mode.
1:23:40Okay, great. Um I'm a data scientist
1:23:42with the I connect team at the World
1:23:45Bank and we've talked about a few
1:23:46different innovative sources of
1:23:48innovative data for road safety. Um so
1:23:52first across many of the contexts that
1:23:54we work with administrative data is
1:23:55often recorded in paper records and our
1:23:57team has used AI and kind of machine
1:24:00learning to help translate uh kind of
1:24:03stacks of paper records into digitized
1:24:06geospatial data sets. So extracting key
1:24:08variables of the time, date, location,
1:24:11types of vehicles, etc., etc. involved
1:24:14in the crash. And we've also worked with
1:24:17police departments as well to think
1:24:18about instead of going from paper
1:24:20records, can we pilot systems so they
1:24:22enter data digitally in in the first
1:24:24place as well. So you start with kind of
1:24:26like nice rich data sets of
1:24:27administrative records to of of crashes
1:24:30to inform um safety work. Um going
1:24:34beyond crashes at individual locations,
1:24:36crashes are even though they're a big
1:24:39problem across cities or rare events at
1:24:41individual locations. So we can think
1:24:42about detecting near mississ crashes
1:24:45from networks of CCTV data across
1:24:47cities. So this is something that we've
1:24:49piloted um at Ababa using kind of their
1:24:52network of CCTV datas using uh computer
1:24:55vision models to to do this. Um, beyond
1:24:59measurement, telematics have often used
1:25:01to kind of monitor vehicles. But we had
1:25:02a team implement a randomized control
1:25:04[clears throat] trial to test providing
1:25:07uh speed and safe driving detected from
1:25:10telematics back to informal transit
1:25:13vehicle owners to see if this would help
1:25:15induce uh their drivers to drive more
1:25:17safely. They found that data alone uh
1:25:21didn't induce them to drive more safely,
1:25:22but telematics in combination with
1:25:24financial incentives did encourage
1:25:26informal transit drivers to to drive
1:25:28more safely. And we can think about um
1:25:32safety from a broader dimension as well.
1:25:35So, we've done lots of work focused on
1:25:36on on road traffic crashes, but also um
1:25:40harassment um or gender- based violence
1:25:43and sexual harassment is particularly
1:25:44prevent prevalent in public transport
1:25:46and kind of prevent um people from
1:25:48taking public transport in in the first
1:25:50place. So, a team member developed a
1:25:53mobile app in Dar Salam um to kind of
1:25:55crowdsource the measurement of sexual
1:25:57harassment across locations including in
1:26:00public transport stations and public
1:26:02transport itself. Um and one of the
1:26:05results is that on average women face an
1:26:07incident of harassment every two weeks
1:26:09and 90% of incidents are occur in public
1:26:11transport. So again this is an issue
1:26:13that's particularly prevalent um in in
1:26:15in the transport sector specifically. Um
1:26:19and again thinking about safety from a
1:26:21bit broader dimension. Um so again those
1:26:24are just some super quick uh use cases.
1:26:26We'll kind of leave it to the discussion
1:26:27to dive in more but there's a QR code if
1:26:29you're interested in I connect work uh
1:26:31more more broadly. Thank you so much.
1:26:34>> Thank you so much, Rob. I'll hand over
1:26:36to Ping. And I see you're in the same
1:26:38room as Liz, but uh starting off with
1:26:39you.
1:26:42>> Good morning, good afternoon, and good
1:26:43evening to everyone. Um my name is Ping
1:26:46and I'm I'm very lucky to be um next to
1:26:48Liz today. I usually am in another
1:26:51country. Uh my work is based in Oakland
1:26:54in New Zealand. And that's the case
1:26:56study that I wanted to share today. Um
1:26:59Oakland is a large local government
1:27:01area. It covers about 7,000 kilometers
1:27:04of roads. Uh since about 2020, we have
1:27:08set um safe speed limits on uh around
1:27:121,500 kilometers of that road. Um that
1:27:16journey has um not at all been smooth
1:27:19nor linear um and has gone through many
1:27:22changes um and and impacts. Um but I
1:27:26think my kind of key story today is that
1:27:28certainly data and the innovative uh use
1:27:30of data has been um a real foundation
1:27:34that has enabled that in innovation and
1:27:36implementation and I'm very grateful to
1:27:39work with some really great people some
1:27:41really smart um data and technology and
1:27:44people who have have done great great
1:27:46engineers and great thinking um who have
1:27:49really led that. So I do want to
1:27:50acknowledge them um and all the work
1:27:52that they've led. Um what I wanted to
1:27:54share today is particularly how valuable
1:27:57data is in various stages of the process
1:28:00in implementing and adopting speeds just
1:28:03to give you an example of where there's
1:28:05real real opportunity for data um and
1:28:08the use of AI including um to make a
1:28:10real difference. So certainly in the
1:28:12proposal stage um what you know there's
1:28:15been some really really fabulous
1:28:17presentations today and so thinking
1:28:19about where to start where to prioritize
1:28:22um those priority scores I've seen or
1:28:24where there's the most benefit um in
1:28:27Oakland we have a version under the IRA
1:28:30family adopted for New Zealand and also
1:28:33for an urban context um and so certainly
1:28:35looking at at that crash risk was a big
1:28:38part of our story looking at data around
1:28:41current operating speeds. Um but for
1:28:43some of the 30 kilometer an hour areas
1:28:46it it wasn't necessarily always um data
1:28:50driven. For example, we started in the
1:28:53city center. So you know the city center
1:28:57made a lot of uh sense from a kind of
1:29:00network planning perspective, a place
1:29:02that people understood, a place that
1:29:04supported walking and cycling. some
1:29:07speed um 30k areas came because they
1:29:10were areas that the community demanded
1:29:12them or decision makers really wanted
1:29:15lower speeds for example in their local
1:29:17town center or shopping area. Um so the
1:29:20the prioritization method does involve a
1:29:23a combination of um kind of crash you
1:29:26know crash risk data as well as things
1:29:29like community input and and decision
1:29:31maker input as well. Um so that that
1:29:34initial phase is obviously really
1:29:36important and that proposal phase and
1:29:38data can play a real role there. Um data
1:29:41is also very helpful during assessments.
1:29:44Um under our legislation at various
1:29:46times we have been required to do road
1:29:48segment by road segment assessments
1:29:50looking at a number of risk safety and
1:29:52travel time uh impacts uh per segment
1:29:56and being such a data heavy and
1:29:58laborious process. Um certainly now uh I
1:30:02can see with AI there's a huge
1:30:04opportunity to do that. This is work
1:30:05that we did we did some years ago. Um
1:30:09one of the most impactful parts of the
1:30:12process in my mind is the use of data
1:30:16and the innovative use of data to help
1:30:18grow constructive public conversations.
1:30:21uh speed change conversations and the
1:30:23delivery of those are by their nature
1:30:25something that attracts a very diverse
1:30:28set of views in the community and that
1:30:31can make implementation very very
1:30:33challenging. Um and I think really good
1:30:36data particularly around what might be
1:30:38some of the travel time impacts can be
1:30:41very very powerful and if that can be
1:30:43provided at a granular level um that can
1:30:47really make a difference when people
1:30:49start to get concerned about proposals.
1:30:51Um and certainly for us in the city
1:30:54center uh it's a it's a it's our port so
1:30:57it's a big freight use. It's obviously
1:30:59the core of much of the business and
1:31:01economy in the region. Um that was a
1:31:04really key area where data played a key
1:31:06role in trying to assess what could be
1:31:08the impacts on the network from changing
1:31:10the speed limit. Um and then and then
1:31:14after implementation during monitoring
1:31:16and evaluation. So again data and and
1:31:20innovation there fundamental role. Um
1:31:23the most recent evaluation we've had
1:31:25looked at a mixture of data sources,
1:31:28operating speed data changes, um
1:31:30crashes, kind of crash crash areas
1:31:33crashes and their um and how they've
1:31:35changed and also hospitalization
1:31:38research based on hospitalization data.
1:31:40And so being able to get a better and
1:31:42more fuller picture particularly of harm
1:31:45for people walking, cycling, and
1:31:47motorcycleycling. that final most recent
1:31:50evaluation showed a reduction in deaths
1:31:52and serious injuries of between 16 and
1:31:5521% depending on the methods used. Um so
1:31:59and and we're we're currently looking at
1:32:01at refreshing that now we have five
1:32:03years of data. Um there is also customer
1:32:07data and then customer data can be quite
1:32:10um quite helpful and pivotal. So um so
1:32:13thank you. I think I just wanted to kind
1:32:14of emphasize how fundamental data can be
1:32:17in all the phases of implementation.
1:32:20>> Super. Thank you very much. So Liz,
1:32:25>> thank you and it's lovely being here. Um
1:32:28I've got a couple of slides. I wasn't
1:32:30sure if I'd show them or not, but I
1:32:32think I might. uh entire screen
1:32:39up
1:32:44just to say a bit of context about
1:32:46Transurban. Um you you mentioned a bit
1:32:49about us before, but we've been
1:32:51delivering and operating Rose for about
1:32:53three decades and there's 22 assets
1:32:55under operation. We're in
1:32:57>> Just quickly want to jump in. Liz, is it
1:32:59just me who's having difficulty with
1:33:01your sound?
1:33:06That's the sound. Can you hear me now?
1:33:11Maybe someone else from
1:33:13>> Yeah, it's very like crackly.
1:33:15>> Is also crackly. Yeah. Mhm.
1:33:21>> Ah, okay.
1:33:24Hold on a moment.
1:33:27I will stop sharing.
1:33:36How do I stop sharing?
1:33:38>> You're not sharing.
1:33:38>> I'm not sharing now. This is that
1:33:40better. I don't share. I'll just talk.
1:33:42>> Yeah, it's a lot better.
1:33:44>> I won't I won't share. So, that's fine.
1:33:47I won't share. I'll just give you some
1:33:49um
1:33:49>> Don't be sorry. Don't be sorry. So, I
1:33:52was saying that we've been operating um
1:33:53roads for about three decades and
1:33:55there's 22 assets under our management.
1:33:59Um they're all high-speed multi-lane
1:34:01motorways that are told. Um we have uh
1:34:052.6 million trips every day. Um and we
1:34:09respond to about,00 incidents per week.
1:34:13Um and we have close to 12 million
1:34:15customers using our network.
1:34:17[clears throat] So it's um it's big
1:34:19volume. Um, and so because of that and
1:34:22that, excuse me, and this roadside
1:34:24technology that we use to manage the
1:34:26network, we have billions of points of
1:34:29data that come into our um, through our
1:34:31control rooms. And with that data, we're
1:34:34starting to apply AI to that. And we've
1:34:36been doing that for quite a while. And
1:34:38that could be everything from um, using
1:34:41predictive analytics to understand where
1:34:44we should position our uh, incident
1:34:46response. So, we offer free incident
1:34:50response to, well, it's not free because
1:34:52it's a toll road, but anybody who breaks
1:34:54down or has an incident on our network,
1:34:57we get to them really quickly within a
1:35:00few minutes and clear them and make the
1:35:02environment safe. And in doing that, we
1:35:05drop the speed, we close the lane, we
1:35:07send out protection for um the motorist
1:35:10who's on the side of the road and also
1:35:12for the crew who is attending to them
1:35:14because roadside and secondary incidents
1:35:16on high-speed roads are incredibly risky
1:35:18as as you can imagine. So when we bring
1:35:21this data in, we can start to predict
1:35:23what we should be doing, where we should
1:35:25be positioning that incident response.
1:35:28And that actually helps um not just
1:35:29safety but also throughput um and gets
1:35:32the network moving again because our
1:35:33corridors act you know very favorably to
1:35:36the freight task as well. So we don't
1:35:38want to be holding up productivity.
1:35:40So um that's that's one example of how
1:35:43we're starting to use it. Another one is
1:35:46um or another couple that I'll just run
1:35:47through is that we do um major
1:35:51maintenance across the life cycle. So
1:35:53that means that we're doing major
1:35:54re-shooting or barrier upgrades or even
1:35:57expansion of the road going from two
1:35:59lanes to three lanes and the like. And
1:36:01that happens across our network all the
1:36:04time on one of the roads that we're
1:36:05operating in. So Ping and I are in
1:36:08Sydney at the moment. We're at the
1:36:09Australasian Road Safety Conference.
1:36:10That's why we're in the same room. Um
1:36:12but we have roads in Melbourne, Sydney,
1:36:13and Brisbane here. Um and so when we're
1:36:17we what we know is that crashes increase
1:36:20during construction. um that's just a
1:36:23factor. It happens across all networks
1:36:25and we're able to we've been using um
1:36:27building dashboards around the speed
1:36:29speed profile of um the people driving
1:36:33through those areas. So regardless of
1:36:35the um the safety barriers that you put
1:36:37in place, the closing lanes, the reduced
1:36:40speed, we still monitor um and find that
1:36:43people are driving well above the speed
1:36:45limit and that will be contributing to
1:36:46crashes. So with these dashboards we're
1:36:49able to share with enforcement agencies
1:36:51and they will help us in terms of
1:36:53operating um or running enforcement
1:36:55campaigns. It helps um us to advocate uh
1:36:59for the right agencies to put in uh
1:37:02automated speed enforcement during those
1:37:05um uh that construction. Um and it also
1:37:08means that we can adjust the um the
1:37:11environment to better suit or to give
1:37:14the right cues to the motorists that
1:37:17you're entering into um a high-risisk
1:37:19environment. So we're able to use our
1:37:22data to start doing that. And what we
1:37:24see will happen into the future is it'll
1:37:26just become automated. So we'll be able
1:37:29to put algorithms into all of that
1:37:30roadside technology and eventually it
1:37:33will be that technology talking to the
1:37:35driver in the vehicle. So that's where
1:37:37we're we're aiming to go. And the last
1:37:39one is you know looking at the
1:37:41operations right now. So we have um a
1:37:43lot of congestion when you saw that that
1:37:45or I mentioned the data we have before
1:37:48you know 11 12 million customers and
1:37:50those you know millions of trips every
1:37:52day during peg times we get um queuing
1:37:55on our network and queuing when they
1:37:57they're leaving the network and there's
1:37:58only so much space. So we've been able
1:38:01to link up some of that roadside
1:38:02technology in terms of automatic
1:38:04incident detection. uh variable speed
1:38:07limit signs, um ve uh variable message
1:38:10signs, um and start to um implement uh a
1:38:15traffic management program that warns um
1:38:18motorists downream um or upstream that
1:38:22they're approaching queuing or
1:38:25congestion. the speed drops um and then
1:38:29we get better throughput and we reduce
1:38:31the rear end crashes which comprise
1:38:33about 50% of the crashes on our network
1:38:36and then we also have reduced um uh
1:38:40incidents where um people get frustrated
1:38:42and dart into the free flow traffic so
1:38:45into the running lanes u which is highly
1:38:47dangerous again. So, we're using our
1:38:50data to be able to start to adjust how
1:38:53the how the the network operates um and
1:38:56then start to feed that into algorithms
1:38:58through our control rooms that will
1:39:00automate um this onto the network. So,
1:39:03data is really powerful and we're very
1:39:05fortunate to have um incredible amounts
1:39:07of data and then also integrate that
1:39:09with third party data like connected
1:39:11vehicle data, weather data. Everyone's
1:39:13talked about IRA today. We've done our
1:39:16rows at IRA data as in an API into a
1:39:19platform that we have and I heard um one
1:39:21of the jurists earlier talk about AIAP
1:39:24and that's something that we're starting
1:39:26to to do as well in building digital
1:39:28twins um through LAR of our network as
1:39:30well. So that's all input into our life
1:39:32cycle management.
1:39:34>> So that's a super interesting.
1:39:36>> Thank you so much all of you for for
1:39:38sharing those very compelling use cases.
1:39:40So I I'll jump right into the to the
1:39:42discussion and I'll remind anyone who
1:39:44wants to share or ask a question to to
1:39:46just type it in the Q&A box. Um I'll
1:39:49start off with you Rob. Um sort of
1:39:51impact evaluations can be seen quite as
1:39:53an academic exercise but you've been
1:39:55really applying them directly on ground.
1:39:57Uh how can transport authorities use
1:39:59these impact evaluations to promote
1:40:01safer roads?
1:40:05>> Yeah, I think in a couple ways. I mean,
1:40:06one, I think it's important to kind of
1:40:08like test our assumptions when we do
1:40:09interventions to kind of make sure
1:40:11they're kind of having the impact that
1:40:12we that we want. Um, but I think a
1:40:15critical thing for road safety as well
1:40:17is that often, especially in like
1:40:19cities, certain road interventions can
1:40:21sometimes be political. Um, so if you're
1:40:24thinking about kind of like
1:40:24pedestrianizing like a certain road,
1:40:26there often might be push back that
1:40:28while it might be good for safety, it's
1:40:30going to cause lots of congestion
1:40:32concerns. So they famously this happened
1:40:35in New York City when they were
1:40:36pedestrian time square. They were going
1:40:38to could have reduced lots of space for
1:40:40cars. Got lots of push back like hey
1:40:42this is going to cause like enormous
1:40:43congestion. But they use some like
1:40:45really nice kind of like telematics and
1:40:47crash data as well to show that number
1:40:49one when they kind of did a temporary
1:40:52pedestrianation kind of programming. Um
1:40:55they found that number one it did reduce
1:40:57crashes but also ended up making speeds
1:41:00more or less go faster just because like
1:41:02the geometry of the road got got
1:41:03simpler. So that was kind of a nice
1:41:05application to kind of like help
1:41:07alleviate concerns when we're doing
1:41:08these kind of like safety efforts as
1:41:10well. um then you kind of think about
1:41:12using these like temporary solutions
1:41:14then making them them more permanent. So
1:41:16I think kind of help kind of alleviate
1:41:18these maybe more kind of political
1:41:20backlashes against or I don't know these
1:41:22kind of like larger concerns about the
1:41:24their balance of like we need to make
1:41:26sure people can like vehicles can travel
1:41:28on their roads but also balancing the uh
1:41:31the safety concerns as well. So I think
1:41:32that's kind of like a key area where
1:41:34impact can impact valuations can can
1:41:37particularly support.
1:41:40>> Thank you. And that that makes me sort
1:41:41of think of ping now because you talked
1:41:43about speed management also u and often
1:41:47can face sort of like a strong sort of
1:41:49public push back a bit to what Rob what
1:41:51was uh sort of a bit saying but uh how
1:41:54did you use data and predictive modeling
1:41:55to sort of help shift this conversation
1:41:57to actually build a community and
1:41:58political buying behind this?
1:42:02Yeah, I think the day that I it really
1:42:04dawned on me how powerful data can be to
1:42:08shift public conversations was the day
1:42:10we did a media briefing about a big
1:42:12proposal of 30k areas and we had had a
1:42:15room of journalists for an hour and a
1:42:17half. We had we had global experts like
1:42:20Dave Cliff from the global road safety
1:42:22partnership. We talked about the safe
1:42:24system. We talked about survivability.
1:42:26We talked about so many benefits. Um,
1:42:28and when the headlines came out, the
1:42:31headlines all referred to a single piece
1:42:33of modeling. And the headline says,
1:42:36"This proposal is expected to cost
1:42:39motorists seconds, not minutes." Um, and
1:42:42it was a a really lovely piece of data
1:42:44and modeling where they'd looked at a
1:42:46number of scenarios, um, traffic
1:42:48modeling. They predicted them for the
1:42:50average um, car trip, which is about a
1:42:5420 minute car drive. and they worked out
1:42:56that regardless of which option decision
1:42:59makers chose, all of the options sat
1:43:02between quite a narrow band of travel
1:43:04time impact. Um, and I think that really
1:43:07brought home to me how a very timely,
1:43:10very well done piece of data and
1:43:11modeling can really turn people um, who
1:43:14maybe previously feared something would
1:43:17have a huge impact and just help them
1:43:19feel like, oh, maybe that will be okay.
1:43:24Thank you so much for for for sharing
1:43:25that. So I mean what we've talked about
1:43:27now is also from both cases here uh
1:43:29using a lot amount of different data to
1:43:31actually model sort of an outcome and
1:43:33and and use that to to to drive change.
1:43:36But I mean Liz, coming back to you, I
1:43:38can only imagine the amount of massive
1:43:40different types of data streams uh
1:43:42coming in from from you and what has
1:43:45been the biggest challenge for you to
1:43:47actually streamline this and and move
1:43:48from a traditional sort of reactive
1:43:50analysis to to predictive analysis.
1:43:54>> It is challenging. I think the quality
1:43:55and integrity of the data is number one.
1:43:58So you know you do need to test that and
1:44:00and data has you know a whole lot of
1:44:02holes and when you operate the number of
1:44:04roads that we have and the variability
1:44:06across the control systems and how
1:44:08people code that data you're on on a
1:44:10pathway on obviously continuous
1:44:12improvement. So a lot of um assurance is
1:44:15required. Um and some things you don't
1:44:17need to have pure data. You know that's
1:44:18enough to get the insights to be able to
1:44:21give an indication across a larger part
1:44:23of the network. Say for instance um
1:44:26where you might be looking at um at at
1:44:29managing speed enough you've got enough
1:44:32data to actually give motorists the
1:44:34right advice at the right time as to why
1:44:38that speed might be dynamically up or
1:44:40down that's okay but when you're looking
1:44:42at something that might require absolute
1:44:44accuracy then that's a different matter.
1:44:46So that that's challenging and then to
1:44:48be able to get the people across our
1:44:50network and who we engage with you know
1:44:52and our customers of course um and the
1:44:55the key one is that they have to have
1:44:57confidence in what we're telling them
1:44:59and that's the key thing in terms of of
1:45:01data and that's and you know AI is going
1:45:03to help that once you the experts you do
1:45:05need expertise in this you just can't
1:45:07throw data into you know a big mixer and
1:45:10say come out with a solution because you
1:45:12have to have the expertise to ask the
1:45:14right questions and if you're going to
1:45:16find those holes, you still have to ask
1:45:17the right questions. So, I think that's
1:45:19the key thing here when we're looking at
1:45:21how we're using this vast, you know,
1:45:24volume of data. Um, it's all important,
1:45:27but some is more important than others.
1:45:32>> And and we're getting close to to the
1:45:34end, so I'll ask a final set of
1:45:36questions. So, I appreciate a quick
1:45:37reply, but a data set I haven't
1:45:39previously thought about in this sort of
1:45:42aspect is crowdsource data. uh Rob uh
1:45:45and like how you've used it I've
1:45:47understood in the Nijani app uh can you
1:45:49maybe give an example of how this can be
1:45:51used to complement sort of
1:45:52administrative or or private sector data
1:45:54to actually inform road safety efforts?
1:45:58>> Yeah, I think a key value of the crowd
1:46:00sourcing data can it can help provide
1:46:03context to
1:46:05uh it kind of helps eliminate things
1:46:08that might be kind of like tricky to
1:46:09capture it through other data sources.
1:46:11So to I mean kind of capture
1:46:14you know where in a city where sexual
1:46:15harassment is kind of most um an issue
1:46:18it kind of relies on it it's helpful to
1:46:21kind of like rely on the crowd to kind
1:46:23of like you know report that and
1:46:25indicate mapping of that. Um another
1:46:28example of how we use crowd sourcing is
1:46:30that in in formal transit in Kenya which
1:46:33typically experiences high rates of of
1:46:35crashes um I mean their safety is often
1:46:38measured through kind of like telematics
1:46:40data. you kind of track their speeds. We
1:46:41also created an app to enable passengers
1:46:44to kind of provide feedback on the
1:46:45driving performance just through kind of
1:46:47like a textbased platform. Um, we got
1:46:50lots of like written descriptions which
1:46:51was nice that provided kind of nice
1:46:53context into why a vehicle might be
1:46:55driven unsafeely such as incidents of
1:46:58vehicle like picking up passengers where
1:47:00it might not be safe like the driver
1:47:02seeming kind of like tired or fatigued.
1:47:05So it kind of provides nice I think kind
1:47:07of context and especially with like that
1:47:09text data now AI is really good at kind
1:47:11of like summarizing that as well. So I
1:47:13think AI can kind of help make better
1:47:15use of of this sort of data as well. Um
1:47:18yeah so I think crowd sourcing useful in
1:47:21many different kind of kinds of
1:47:22applications
1:47:25>> and and thank you so much and I think uh
1:47:27we have two minutes left so I'm going to
1:47:29wrap up our session. uh so many
1:47:32interesting topics and it felt like we
1:47:34needed another good 30 minutes to to
1:47:36discuss this even more uh but thank you
1:47:38so much uh for sharing these insights. I
1:47:41think the the clear thing for me today
1:47:43is is how data is really used today.
1:47:45It's not just analytical tool but
1:47:47they're active sort of safety mechanisms
1:47:49on ground driving change. So I think
1:47:51from Rob I mean understanding how you're
1:47:53combining crowdsource data with other
1:47:55sort of regular evaluation uh efforts
1:47:58help sort of target these safety
1:48:00interventions. So, so thank you very
1:48:02much. And and Ping, super interesting to
1:48:04see how your modeling has really helped
1:48:06build public trust and and streamline
1:48:08sort of evident based data for speed
1:48:11management. And and Liz, thank you so
1:48:12much for for sharing how you've been
1:48:14integrating all these various different
1:48:16data sets uh from roadside technologies
1:48:19and telemetry to really help improve
1:48:21complex environments and and protect
1:48:23highway uh users in real time. So from
1:48:26my perspective, I mean, I'm a true
1:48:28believer in predictive analytics uh as
1:48:30the fastest way to really prevent
1:48:32crashes before they happen. And I think
1:48:34that today's session has really proved
1:48:36that when we use sort of technology and
1:48:38and humanentric design all together, we
1:48:40really help move closer to the to the
1:48:42lovely vision zero target. So thank you
1:48:45so much uh all of you.
1:48:48>> Thank you. Thank you.
1:48:55Thank you Johanna and thank you for the
1:48:58panelist
1:48:59for the very insightful panel. Our jury
1:49:03has returned. Welcome back. Um before we
1:49:08hear from them, um let's settle one
1:49:12thing first. Um the award
1:49:17that you all decided. So, let's go into
1:49:20the results section. Very exciting. Oh
1:49:24my gosh, I'm I'm sweating from
1:49:28excitement here. I don't know if you
1:49:30are. Um, but before the panel
1:49:33discussion, you've been voting for the
1:49:35solution you thought was the most
1:49:37impactful. And the votes are in. Um, can
1:49:42we have a drum roll, please?
1:49:48This team received 38% of all the votes
1:49:52today. Uh the audience choice award goes
1:49:56to
1:50:02team five. Uh congratulations
1:50:09>> audience has spoken and clearly thought
1:50:13your solution would deliver impact.
1:50:16Congratulations. Well done.
1:50:19>> Thank you.
1:50:24Now, um, moving on.
1:50:29Uh now to our jury
1:50:32and I would like to uh put you on a spot
1:50:36Richard since you have been on this ride
1:50:38with us since the beginning as our
1:50:41technical expert
1:50:43and you have spent a lot of time with
1:50:45these solutions.
1:50:46How would you reflect on this journey uh
1:50:50from the applications to the refinement
1:50:53phase to the of today?
1:50:56Uh yeah, I'd like to congratulate
1:50:58everybody again today for their hard
1:51:00work and their pitches. They were
1:51:02incredibly well done. Um you're you're
1:51:05all winners in in my eyes and I think
1:51:07you know there was an enormous amount of
1:51:08competition just to get through to this
1:51:10phase. You know we had something in the
1:51:11region of 40 different applicants
1:51:14originally. Um and you know you were
1:51:17selected because you know your solutions
1:51:19stood out. they were doing something
1:51:21different or they were taking a more
1:51:23common approach and applying a level of
1:51:25robustness to it. Um what was also great
1:51:28was seeing how um everybody um refined
1:51:32their solutions, took on board the
1:51:33comments from those first phases um and
1:51:36adjusted them and then also the way that
1:51:39they um used the EZRI solution as well.
1:51:42Um most people used um other um uh tools
1:51:46and solutions to rapidly scale a
1:51:48demonstration. Then thinking about how
1:51:51they took that and then developed it
1:51:53into more of an enterprise solution was
1:51:56was really fantastic. Um the last thing
1:51:58I'd like to just comment on briefly um
1:52:00is is around AI and the use of it within
1:52:03this challenge. I mean this is this is
1:52:04something where we're trying to use AI
1:52:08and and people are using um different
1:52:10kinds of AI solutions. I just caught the
1:52:13end of the the presentation and listened
1:52:15to some of the things that Marty was
1:52:16saying at the end there. you know, um AI
1:52:19is really powerful. It's moving on an
1:52:22absolutely rapid pace, and we need to
1:52:24make sure that um we're using the right
1:52:27AI, but we're also making sure that
1:52:29we're doing those checks as well. And I
1:52:31think this really um emphasizes the need
1:52:33for having the experts in the room. You
1:52:36know, if you're not an expert and you're
1:52:37using AI to deliver solutions, they it
1:52:40can come across as as quite weak and it
1:52:42and it wouldn't show your level of
1:52:44understanding. So making sure that you
1:52:46understand that policy policy context
1:52:48you use the research love the way that
1:52:50people quoted a lot of the research um
1:52:52within that that really grounds the the
1:52:55AI and the tools that you've developed
1:52:57in the literature in the evidence in the
1:53:00policy goals as well. So yeah, a really
1:53:02clear message to everybody who's uh
1:53:03hopefully still listening to this is,
1:53:05you know, make sure that you are
1:53:07checking, make sure that you're um uh
1:53:09getting your AI outputs, the AI outputs
1:53:11that align with, you know, the um the
1:53:14the facts and and everything that we
1:53:16already know about what can be done to
1:53:18improve um road safety, save lives um by
1:53:22setting uh safe speed limits.
1:53:26Thank you so much, Richard, for all that
1:53:29you have put in this making this
1:53:31challenge a success. Thank you so much.
1:53:34Now,
1:53:36next we have uh the second place, the
1:53:39runner up and um
1:53:43is everybody ready?
1:53:45the moment we have been waiting for and
1:53:48the jury to um announce the second
1:53:51place. Uh now before um
1:53:56Dimma will be doing the announcement and
1:53:59before you reveal the runner up, could
1:54:01you tell us um in couple of sentences
1:54:04what distinguished this solution and
1:54:07then reveal the name of the challenge
1:54:09and we'll get the drum roll.
1:54:13>> Sure.
1:54:14Okay. So
1:54:17I think two things made the difference
1:54:19that's the scalability and usability and
1:54:22uh I see everybody has done very good
1:54:25job uh but of course there is a little
1:54:28difference here and there and uh based
1:54:30on that we are going to announce uh
1:54:32runner up. So yeah here is let me take
1:54:37the web page.
1:54:41Yeah.
1:54:43Okay. So, yeah. So, yeah, I can see on
1:54:47behalf of the jury, it's my pleasure to
1:54:49announce our runner up. That's um yeah,
1:54:53Neural City. Um I think they built the
1:54:56most complete decision support tool of
1:54:59the four. Uh pick any road and the
1:55:02planner can see why it scored the way it
1:55:04did, what to do about it, how many
1:55:06casualities and action could prevent.
1:55:09So, it was a kind of a complete u map to
1:55:12a decision. Uh the team also looked at
1:55:14junction hotspot and motorcyclist on
1:55:17their own. Uh covered both Thailand and
1:55:19Maharashtra which was a good one and was
1:55:22upfront about that the tool can and
1:55:24cannot do. That's uh yeah honestly uh
1:55:26builds the trust as well. Uh many
1:55:29congratulations to the team. Way to go.
1:55:32>> Thank you members. Thank you everyone.
1:55:36>> Thank you so much Sema for that and
1:55:39congratulations Neural City.
1:55:42Yes, thank you.
1:55:44>> And now um let's move into the first
1:55:48place,
1:55:50the moment that we have all been waiting
1:55:53for. And here to announce we have Holly.
1:55:58Will you do us the honors and first tell
1:56:00us why the jury sources team and finally
1:56:04make the announcement?
1:56:06>> Thank you. Uh this is quite an honor. Uh
1:56:10first I have to say it was a very
1:56:12difficult decision. Uh a lot of
1:56:15discussion amongst the jurors. All four
1:56:18teams brought something unique and
1:56:20interesting and new and I and kudos to
1:56:24the organizers of the event for being
1:56:26able to surface all these wonderful
1:56:28starts into future research and
1:56:30applications. Uh the team we selected uh
1:56:34we felt truly understood the challenge
1:56:38and we know how difficult it is to turn
1:56:41a model into a user interface and we
1:56:45felt this team not only created an
1:56:48interface that was intuitive. Uh but
1:56:51when they presented their scores that
1:56:54came from the model, they were
1:56:55accompanied by a clear explanation of
1:56:58where those scores came from which is so
1:57:00valuable. Our judges felt this solution
1:57:03was scalable and the judges noted the
1:57:06attention this team paid to detail uh
1:57:08such as not to give anything away, but
1:57:11the team did add features such as uh
1:57:15activity noting activity around schools,
1:57:18points of interest to help identify
1:57:20areas where pedestrians may concentrate.
1:57:24And finally, our judges felt that this
1:57:27application had room to grow. they were
1:57:30already excited to add comments on how
1:57:32it could improve. So without further
1:57:35ado, uh we would like to congratulate 5S
1:57:38uh on [clears throat] winning first
1:57:40place uh in the competition.
1:57:46Congrats.
1:57:48Congratulations
1:57:50team 5s
1:57:52uh the winner of the AI for safer road
1:57:55safer speech innovation challenge 20126
1:58:01and uh you have been selected for the
1:58:05opportunity to take this further with
1:58:07ADB's transport team. Now, team 5S, uh,
1:58:13would you like to make a small, um, say
1:58:17say a few words?
1:58:19>> Mhm.
1:58:21Uh, thank you for this opportunity and
1:58:23thank you for the everyone who came
1:58:25today and I'm really honored to see this
1:58:29result and
1:58:31uh, I used to be a banker and I changed
1:58:33my career to traffic engineering uh,
1:58:36because I wanted to save lives. So this
1:58:38has been my dream job uh since uh 2019
1:58:42or 2017. So I really hope I'm really
1:58:45hoping we are improving our products and
1:58:48uh polish polish it and actually save
1:58:50lives through this. VC do you have?
1:58:54>> Yeah, thank you for uh the team uh uh
1:58:58Julie members and all the uh
1:59:01coordination team, technology team, my
1:59:04other competitors uh all presentation
1:59:06were really nice. The agenda behind this
1:59:10behind this whole solution was we have
1:59:12been hearing news about some of our
1:59:15relatives, some of some of our close
1:59:17ones getting hurt, getting killed in an
1:59:21crash and we being transportation
1:59:23engineers, I think this was a perfect
1:59:25opportunity to work this solution in a
1:59:28developing country and I belong from a
1:59:30developing country like India. So, uh
1:59:33really thank you for the opportunity.
1:59:36Congratulations
1:59:41once more to to the winners to the
1:59:44runner up um to all our finalist teams
1:59:49reaching this stage out of 466
1:59:52submissions that is an achievement in
1:59:55itself and everyone here should be proud
1:59:59for what they have built.
2:00:02Now this is not the end of the road.
2:00:05today's selected solution will now have
2:00:07the opportunity to work together with
2:00:08Asian Development Bank's uh transport
2:00:11sector and information technology
2:00:13department to take their solution
2:00:14further. to briefly show us how your
2:00:18solution will be integrated into the
2:00:20ADB's larger GIS system uh that will
2:00:23inform ADBY projects. I would like to
2:00:27invite my colleague Luthther Linda
2:00:29spatial data infrastructure specialist
2:00:32at ADB to say a few words.
2:00:36>> Thank you very much TA and very excited
2:00:39to see all these products. maybe to
2:00:40highlight um while the slides are being
2:00:43pulled up the two slides that actually
2:00:45everyone that's been on this challenge
2:00:47now has already used uh the ADB
2:00:49enterprise GIS sort of the applications
2:00:51that you've seen already deployed on it
2:00:54and uh and I will just quickly say a few
2:00:56words about the um environment and and
2:00:59what we're trying to do. So we're going
2:01:00a bit back from the specific
2:01:02applications to the bigger picture why
2:01:04actually organizations like ADB but
2:01:06others and organizations um should
2:01:08actually or require GIS at all. So the
2:01:11example of ADB and the transport sector
2:01:14we we have a situation where 70% of the
2:01:16portfolio is in is falling into sectors
2:01:19that are intrinsically geospatial. So
2:01:22the the money and the investment goes
2:01:23into hard assets. It's not there's
2:01:25capacity building and soft assets but a
2:01:27lot of this goes into infrastructure on
2:01:28the ground and that is intrinsically
2:01:30geospatial information and if you do
2:01:33want to handle and build a sustainable
2:01:35portfolio and and identify opportunities
2:01:38you do need to employ geospatial data.
2:01:41So the why um is already there. Next
2:01:46the second question is is are
2:01:48organizations already using it and
2:01:49there's often the perception at least
2:01:51for a long time there was a perception
2:01:52that that it is not done and that's not
2:01:54true. ADB is actually using in a very
2:01:57systematic fashion at the project level
2:01:59um GIS and geospatial data. Um we are
2:02:02we've looked be through procurements
2:02:04etc. And we know that a significant
2:02:06number is actually employment
2:02:07technology. There's downsides to it
2:02:09because it's done on a silo level. So
2:02:11every project is starting the whole GS
2:02:14workflow from scratch to the end within
2:02:16this project and it's not connecting
2:02:18with each other. And that puts us to the
2:02:20third um to the third level. How do we
2:02:22bring this together to make this
2:02:24information uh the next step to make
2:02:26this information more efficient and and
2:02:28help the ADB and the project planning
2:02:30workflow and that is to establish an
2:02:32enterpriseg solution that connects
2:02:34between the projects that is taking the
2:02:37information afterwards for efficient uh
2:02:39monitoring and evaluation on the sector
2:02:41level and to feed that back into the
2:02:43next sector level strategizing um
2:02:46basically brings all this together. Next
2:02:48slide.
2:02:49Just as a quick picture to to put this
2:02:52together. Uh um we have a lot of data
2:02:55that uh that's stick to sell on the left
2:02:57side. For instance, internal external
2:02:59data, non-spatial uh at varian scales
2:03:02and standards which are not necessarily
2:03:04coming together easily. Next the next
2:03:06two
2:03:08next and we have a lot of applications
2:03:10that are built outside or within a
2:03:12system. And what what we're trying to
2:03:14establish with the with the enterprise
2:03:16JS solution is this the system in the
2:03:18middle the um the platform in the middle
2:03:20that brings the data in that unifies the
2:03:23data aligns them makes them compliant uh
2:03:26ensures uh quality um quality standards
2:03:30being applied and then basically feed
2:03:32that back to the other to the um client
2:03:34applications which which is what you
2:03:36have seen right now uh the client
2:03:38applications that the teams have
2:03:39developed others and that basically the
2:03:42data and the client applications
2:03:44connecting through one interface that's
2:03:46becoming data and also hosting for
2:03:48applications uh wise the uh single
2:03:51source of truth and that is um turning
2:03:54uh the the silos that are done on the
2:03:57project level into a unified solution
2:03:59that will improve efficiency and and the
2:04:01sustainability
2:04:03um uh of the uh ADB project portfolio.
2:04:08I hope I stayed with my time.
2:04:11You you did great, Lar. Thank you so
2:04:14much. And now before we close, we would
2:04:18actually like to take everybody. Thank
2:04:21you so much for joining us. To close, I
2:04:24would like to extend our deepest
2:04:27gratitude to our jury for their time and
2:04:30expertise, to the speakers, to Johanna
2:04:34and our excellent panelists, to our
2:04:37partners, the Asia-Pacific Road Safety
2:04:40Observatory, the World Bank Group
2:04:42colleagues AI for Good and ITU, ES3 and
2:04:47TomTom, and to the Japan Fund for
2:04:51Prosperous and Resilient Asia. the
2:04:53Pacific and the High Level Technology
2:04:56Fund who whose support made this
2:04:58challenge possible
2:05:01and thank you to the organizing team for
2:05:04your tireless work. Pretty Gautam,
2:05:06Christine Lucero, Daphne Marcelo, Johnny
2:05:10Salang, Ooi, Bla1 Murph, and Mojo
2:05:13Monahan from ADP. Thank you Rob Marty
2:05:17and Daniela Martinez Perez from the
2:05:20World Bank Group for your lovely
2:05:22collaboration
2:05:24and the technical experts Richard from
2:05:27Agillis Larinda from ADB and Vanessa
2:05:32Copnet Wetszel and Ela Okambo from S3
2:05:36and thanks to all of you who join us
2:05:39today in the audience. Uh we will be
2:05:42sharing today's results on the
2:05:44challenges.adb adb.org/ai4safer
2:05:48roads website and over social media. So
2:05:51keep watching this space while um we
2:05:55will also share updates on the on the
2:05:57upcoming pilot. Now it's time to say
2:06:00goodbye. Thank you all for joining. Safe
2:06:04travels and safer roads.

About this transcript

This page contains the full transcript of AI for Safer Roads Final Pitch Event by ADB Digital Sandbox, generated from the public captions YouTube serves with the video. The transcript has 18,434 words across 2,795 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.