Hacker News new | past | comments | ask | show | jobs | submit login
AI predicts accident hot-spots from satellite imagery and GPS data (unite.ai)
181 points by Hard_Space 6 days ago | hide | past | favorite | 76 comments

It’s interesting that they are doing this, but it’s not particularly novel. The vast majority of accidents happen on roads that meet certain criteria:

- Speed Limit > 30mph and < 60 mph

- lots of driveways and left turns

- moderately heavy traffic at some times of day but not constant gridlock

Just picture the commercial drag strip and the collector roads that wind between huge garden apartment complexes or gated subdivision.

Accidents are especially likely where two such streets cross.


These are the futon of transportation. They’re designed first for high speed and volume of traffic, but also high direct access to the surrounding businesses. These objectives are horribly in conflict and not safe to combine. People want to drive FAST on these, and they do, but they are frustrated by the stop and go nature and there is a constant stream of “surprises” when other drivers take a chance on a turn through a tight gap or run a light that just turned red because they’re sick of waiting. One moment of distraction for either party and BAM, a crash.

Good streets for local access are small and slow moving (high volume can be accommodated by a parallel network of such streets, ie “a grid” but it doesn’t have to be perfect squares). Good roads for high speed are wide, clear, and simple, with lots of extra room to cushion mistakes (medians, shoulders, etc.)

All this model has really done is learned to recognize new “stroads” that don’t have a crash history yet, but will soon.

Wow this video perfectly captures what I thought about visiting the US. At the time I couldn't describe what was making me critical of the way it was built, but this is so well described.

In Europe a street as wide as that would be a main thoroughfare. I seemed to see them all over in the US, like every road in a grid would be super wide, but also have shops.

Stroads are really terrible. It's amazing how they can be legal in US.

What's really terrible about stroads is the phenomenon of the later "bypass" roads that are built long after the stroad/city center turns into a an absolute clusterfuck and aren't usable for the kind of high-speed transit they're designed for. The bypass roads usually take the form of a parkway or some other roadtype with limited entrances/exits.

But then later, because the bypass road is usually built on cheap land outside of the city center, all that property gets bought up and developed as well, and the property owners will demand ingress/egress to their properties. Since the environment is nicer than the snarled up city center, people will start to favor businesses and housing in this area. Over the next 10-20 years the bypass will simply turn into another stroad. To make matters worse, the bypass and the original road will often have the same road numbers, differentiated by either "business" or "bypass", making navigation confusing.

Where I live I often see multi-million dollar homes being built right next to high-speed roads, and instead of a feeder street system taking cars to a more limited number of ingress/egress points, they simply dump the driveways out into the 60mph traffic. Total insanity.

It really does all come down to poor planning, and a lack of desire to make street planned city center grids with mixed-zone housing and municipal managed central parking areas. Instead each business sits in an island in a sea of their own giant parking lots, which often sit mostly unused which makes urban centralization impossible.

Then if you live long enough you get to see the bypass get a bypass of its own and the entire process starts over again...

Because the engineering design doesn’t just prioritize cars it fundamentally presumes cars. If the result is something unsafe:

1) they will use individual responsibility arguments on drivers to blame them

2) blame pedestrians or cyclists or whoever for failing to defer to cars

It’s not a legality problem it’s a philosophy one.

> It’s not a legality problem it’s a philosophy one.

This is not a dichotomy. The legal aspect comes from the "philosophical" one that comes for lobby money.

I agree, but I think for the purposes of this conversation it fits...laws emerge from the philosophy. simply making an effort to change the laws won't change the philosophy, it will just be an anachronism standing against the dominant philosophy.

This doesn't fully explain the phenomenon because stroads are terrible for drivers.

They cause congestion and are much slower than actual roads, while still being unfriendly to pedestrians and bikers. They're bad for drivers and bad for pedestrians.

This probably sounds pedantic but…they are terrible for drivers but they are good for cars. They are designed for cars not drivers effectively. They suck at vehicle throughout in a broad sense but they are theoretically more efficient than the other design options available.

The pramaterizarion is largely around the dynamics of the cars not human behavior. It’s the over algorithmization of decision making. It’s like YouTube radicalizing people by showing them videos that increase engagement. It’s not intentional in a strict sense, it’s just mindless adherence to a process without thinking.

They're cheap and they get land owners who want access to shut up. "there, you have your road, now stop telling us how to spend your tax dollars"

It's even worse. They're not cheap at all. They're cheap to build but expensive to maintain.


The "International Fire Code" (A US regulation used nowhere else outside of North America) mandates roads 20 to 26 feet wide for fire equipment access: https://twitter.com/graykimbrough/status/1404824443600490501

One can assume from this requirement that fighting fires is impossible in Europe, and residential buildings are constantly burning to the ground with firefighters standing around, helpless to intervene.

In Japan they simply have small fire trucks and small garbage trucks for small roads.

Aren't US homes wood framed and EU homes more likely to be made from solid brick/rock?

Grenfell Tower in the UK was made out of concrete. Burning exterior cladding killed 72 people: https://en.wikipedia.org/wiki/Grenfell_Tower_fire

depends on where you are in the EU I think, UK and Ireland still use piles of blocks but no shortage of wood houses in Europe. There's a whole industry of building woodframe houses in the Baltics and shipping them around Europe, incidentally.

Brick houses still burn; interior walls, floors, ceilings, and furnishings are still made from wood or other flammable materials. There have been plenty of castles that caught fire, despite their stone walls.

I talked about this a little bit yesterday, but I live in an area with a lot of traffic fatalities and it seems to me to be, at least in part, an economic issue.

I live in one of the poorest neighborhoods on the outskirts of my US city. There are a few major roads headed toward the more affluent neighborhoods and downtown. What I see when I run along/across or drive along those roads is mostly through-traffic. There aren't a lot of "shops" in my neighborhood, but one or two strip-malls and a few gas-stations and fast-food "restaurants".

Anecdata: In my experience, the areas with less through-traffic (and thus high-danger traffic) are areas with destinations, whether they're shops for goods and/or services or restaurants (that don't involve cars). The reason I think it's economic is because it's not just about the customers/consumers visiting these areas, but the workers. I'd bet that a not-insignificant number of workers in my neighborhood drive to another neighborhood for work, but the inverse is probably not true (that a lot of folks from other neighborhoods drive here for work).

I think part of the solution is to make communities smaller and more friendly to their inhabitants. There should be opportunities for work in these poorer neighborhoods, and that means the services to support them.

It doesn't matter what number is on the speed limit sign on these major roads, people will drive 50 anyway (I have been passed in the turn lane several times because I drive the speed limit on a 2-lane (one each direction) + turn-lane road).

On your last point, a friend who works in urban planning talks about the idea of designing roads to encourage a certain speed rather than relying on speed limits. Narrower streets or intrusions for greenery, etc will tend to lower the speed of traffic.

It would be impossible for a non-expert to recognize these roads without hand-holding training. There are few quantitative criteria ("lots of"), and many that are only understood relative to a persons experience ("Just picture ... "). Yet, they are well-defined and the problem is well understood.

Perfect application of AI/ML. The reasoning behind AI/ML does not need to be mysterious for it to be an appropriate solution. "All it has done is <Something I understand> " is not a valid criticism of the solution.

"All it has done is add these values and subtract these values" would not be a valid criticism of a banking application, for example.

I once heard AI described as (paraphrasing), "Imagine you have an army of seven-year-olds that you can instruct to do a simple task for 24 hours a day"

I remember reading somewhere some transit authority was trying to improve traffic flow into either the Lincoln tunnel or Holland tunnel.

The result of the simulation was both interesting but also unrealistic.

In order to improve intake from the feeder streets the simulation recommended something no normal drivers would do in rush hour: IIRC it wanted drivers to obediently and predictably do alternate lane merges as well as some intricate braided flow pattern. It was great if you were working with logical components but utterly impractical in reality. I don’t think any semblance of that system was ever implemented as it was obvious it would fail worse than the current bad design.

Zipper merges! It’s my understanding that that’s a norm in some countries, but that’s a social issue and not an infrastructure one.

Yes but more than just zipper merges. There was street realignment which also included a braided flow (crossing lanes because not only would they need to funnel traffic from so of canal to the tunnel but also allow them to get off on a local street before the point of no return to the tunnel and same for the ‘uptown’ side) anyway it was an over complicated solution that only worked in simulation and not with real cars and trucks so none of it was implemented.

I'm less familiar with downtown but the whole area around where the Lincoln Tunnel access is in Midtown is a snarled mess. You have the Port Authority, aka one of the world's worst bus terminals, and more generally just a bunch of intersections that are such a mess that you need cops directing traffic because otherwise you just get gridlock between cars trying to make turns and pedestrians trying to get across streets.

Most of society seems to have a problem with the notion of allowing someone else first ultimately means faster for everyone vs the asshats that actively speed up to block/deny someone from being able to merge ahead of them. To me, these are the same people that complain about a single longer queue rather than individual queues for each register.

Zipper is best and I completely agree that "society as a whole" just doesn't understand it and are outright hostile towards the correct behaviour.

Whenever traffic is slow, I make a point of always driving up exactly towards the merge and just sliding into the flow of traffic. Many many times, someone will be pissed at that, either on the way there because my lane is free, sometimes for a very very long stretch and they try to cut me off because I'm 'bypassing' the line or even at the merge, where I just let someone through but then it's my turn but they cut me off.

What do these people think? The 2 lane highway is supposed to become a 1 lane highway just because there's a merge ahead (say because of an accident)? How far back is it supposed to become 1 lane? 100 meters? 500 meters? 10 kilometers? I've seen some crazy long lineups with nobody on the second lane for multiple kilometers. The line could easily have been half as long if everyone just stayed on their lane until the merge point and zippered.

I remember seeing an experiment on TV (Germany) like 20 years ago, where they had a large truck adhere to the zipper rule, merging exactly where they're supposed to and filming it (doing it over and over) and there were many many cases of the truck having to brake hard as people tried to 'slip by'. This was a 'soft zipper' i.e. lots of space to the front with just lines marked out where the truck could come to a standstill/slow down and go on, so it was a safe experiment.

Which ultimately is one of those cases that demonstrate that some ideas are good and 'correct' but "not implementable with humans."

These have been implemented in a lot of places extensively, such as Germany. There are also some implementations in the US. One such that I drove on often is when I-70 eastbound runs into the beltway at Baltimore, MD.

The only problem with zipper merging (much like roundabouts) is a lack of education around how they work and WHY they work. Adjust the education we give new drivers and over time our roads will become more efficient.

CGP Gray's simple solution to traffic and good visuals.


The real simple solution to traffic is no more monkeys driving cars. Also, implicitly, fully isolate cars from human powered forms of transportation.

There is a slight flaw there. It fails to account for vehicle malfunctions (blown tires, running out of power, component failure, roadway debris, etc.) so even with computerized coördination you need to keep a speed determined safe distance from vehicles in front of them. (with reference to the accordion effect of stop and go.)

Accidents will require cleanup services.

Meanwhile self-driving cars can learn of them from the scheduling coordination services and distribute the load among working routes without rubbernecking.

People are taught to use turn signals, and are reminded "keep right except to pass" every 2 miles. Unless there are a lot of blatantly illiterate people driving around, they just don't care about the conventions of the road, and I doubt education is going to change that.

Well I think it might actually democratise some of the business logic you mention. Sure it makes sense from an experts view, yet these designs are still made plentiful. So what if you embed this approach in your development plan to have it develop around it. Eg implement it as a cost function in the overall implementation. I think that’s pretty sweet, and it can also validate the statement you made in one fell swoop because it is a model.

In the 70s and 80s it seemed common for these types of situations to be handled by access roads that ran parallel to the through road. The access road handled all of the individual driveways, was lower speed, and had controlled entrances and exits to the through road. Access roads are still being built but they seem to be much shorter and serve few entities than the designs from the 70s and 80s, which results in more intersections between them and the main through road. My best guess for this design change is that businesses further along the access roads were at a disadvantage as they were less convenient to get to than businesses or properties closer to the entrance point.

>These are the futon of transportation.

Not getting the futon analogy?

Futons are neither good couches nor good beds.

Which is not generally true if you have good (i.e. more expensive) futons.

And likewise if you throw a dedicated turn lane and some sidewalks at these roads they do a pretty ok job.

Absolutely not. If you're ever in Calgary, take a walk along Macleod trail near the Chinook centre. The sidewalks are filled with face-height street signage in the walkway, the sidewalks randomly venture between vehicle lanes, they'll sometimes just end, there are pedestrian road crossings without markings or signage, etc. It's one of the most unpleasant built environments I've ever encountered, up there with polluted Soviet mining cities.

I'm sure someone somewhere has managed to do it wrong but where I live it's done decently well and there's a fair number of main roads that meet the description I gave and are fine as a pedestrian, not perfect but fine.

Unprotected lefts have been widely recognized as one of the tasks that are especially difficult for autonomous driving to handle--which, of course, means that they can be relatively tricky for people too. And, yes, well-defined places to walk and cross streets help pedestrians--at least if they use those defined places which they often don't in cities.

What? Futons are not at all couches and great beds (for some people anyway, but I like’em). Do americans mean something weird by futons?

Yes, there was a fad for making mattresses that sort of fold into a weak, uncomfortable couch. That combo is called a futon here. Now they have frames that give them some supportive structure, but the earlier ones looked like this: https://thehousingforum.com/single-futon/

Ah so it's a kind of sofa / pullout bed, with a continuous folding mattress instead of the older "split" style

So is AI really needed here?

Futons work pretty well unless you need a 10/10ths bed or a 10/10ths couch. These roads are the same. They're fine until you load them with enough traffic to make all the intersections dangerous but not enough to slow them down.

They're really not; even outside of their horrendous safety record they are also responsible for the death of pedestrian traffic in the US and canada

To be clear, high-end futons are actually really nice. Stroads have horrible safety records, futons do not. However there is always a trip and fall hazard around furniture. Please exercise caution when maneuvering indoors.

What's your point? Every road with mixed traffic is going to be responsible for the death of pedestrians. If pedestrians were on highways they would get killed there too.

These roads aren't ideal but they're everywhere because they're cheap. Unless you can wave a magic wand and make municipalities rich enough to carpet bomb everything with traffic lights, dedicated turn lanes and sidewalks there are going to be tradeoffs.

And carpet bombing traffic lights everywhere isn't really an answer even aside from cost. When some new retail went in near me, a bunch of traffic lights were added. To be clear, the adjacent interstate exchange was a really lousy intersection at busy times of the day. But the mass of street lights now make a straight shot through the intersections about a 5-10 minute process to go maybe a quarter mile.

>But the mass of street lights now make a straight shot through the intersections about a 5-10 minute process to go maybe a quarter mile.

I'm sure all sorts of wonderful pedestrian friendly moves get pulled by people trying to get ahead in that game.

There are essentially no pedestrians there (and no explicit pedestrian crossing). It's basically the boundary between a mostly rural exurb and a small city's sprawl. What it does have is one of those stupid zipper merges immediately after a traffic light which encourages aggressive drivers to jump the merge queue on a regular basis even if it means effectively running someone off the road.

I highly recommend the linked video (posted just above). It is not a natural law that pedestrians die on streets/roads, but it is a result of this typical North American road design.

I think he meant people don't walk anymore because the roads are fundamentally made for cars.

The fact that you can make out every single historical crash location in their prediction shows that this model is horribly overfit. (Edit: The paper does mention this general problem.)

Also, the rest of the prediction appears to just be the density of the road segmentation. Look at the parking lot (?) top left of the purple box (or all the road-rich neighborhoods in the right half):


Plenty high estimated risk, zero actual collisions.

You could get the same predictive quality with a simple gaussian blur of the "historical accidents" plots (road segmentation included), it seems.

>You could get the same predictive quality with a simple gaussian blur of the "historical accidents" plots (road segmentation included), it seems.

I worked on a project like this for an insurance-oriented hackathon

We created a heatmap from readily available historical data adjusted for traffic density, and played a sound on a mobile device when approaching "hotspots".

For extra "wow" points we actually had someone drive out in my car and tracked them live as they drove around the location of the event, with hotspots showing up at places like where a blind entrance to the parking garage intersected with the road

> This is a step toward general AI, because our model can predict crash maps in uncharted territories.”

I hope they are misquoted here.

The word "general" does indeed not appear in this sense in the paper.

I read the underlying research publication, and linked editorial has overhyped the results beyond all reason.

The main takeaway from this work is that better data equals better models. The data fusion approach taken by the authors is the most interesting thing about it. The comparisons to baselines is the weakest part of the article. The effort to sell this as a highly significant result is just sad, but is mostly a reflection of the state of academic publishing.

It's an interesting problem.

If I am an accident oracle -- say I know with certainty where, say, 50% of accidents will occur (the other 50% are truly stochastic with no structural component at all) -- and the city believes me, then surely when I tell them an accident will occur at Intersection X, they will take measures that prevent accidents at Intersection X.

But this also means any measure of my ability to predict accidents is confounded, because the equilibrium behaviour would be that I never predict an accident and plenty of other accidents occur. Moreover, because of all the other confounders, it's actually unclear whether we should expect accidents to go down or up or stay the same or what, and to the extent year to year variability was already quite high the problem will be even larger, so even high level numbers before and after hiring my services aren't easily interpretable.

Which is fine, you just need to convince decision makers of this particular inferential fallacy and then hopefully they keep listening even though the KPI is wrong. Except what if the type of accidents change and I'm no longer an accident oracle? Then the mitigation efforts are wasted, and also the money they pay me.

One solution would be to basically engage in some kind of RCT where some of my predictions are held out for assessing my model while others are acted on and where the strength of the prior about my effect decays. Good luck telling voters and lawyers that justification, though.

Fun stuff.

If the city does something to reduce the risk of some of the pointed areas, does your oracle reclassify it as lower risk?

If you make an oracle that simply points that some necessary geometries accumulate the risks, don't be surprised when every authority not only ignores you, but also becomes very annoyed if you insist on it.

The real solution is not only to predict but to explain why the accidents happen, there are always some reasons why some places/location have more accidents then normal.

Right? Say if you know there will be a 2% chance to have a 1% increase over the yearly average at LocationX on Thursday…what can you do? Add police and first responders? Reroute traffic? Tell people to stay home? I feel like nothing that can be done will be able to solve the problem but just move it somewhere else

That’s a bit defeatist.

The Dutch took the approach of treating car crashes similarly to how the FAA treats plane crashes: do a deep analysis of what occurred, what the root structural issues were (for instance, is it difficult for a car turning left to see oncoming traffic?) and then change the environment to improve the structural issues.

Knowing where incident hot spots are likely to be should help you redesign the areas most likely to cause issues.

There’s a reason the Netherlands has around 1/3 as many traffic fatalities per capita as the US: in the US, it seems like we throw up our hands at problems like road crashes, whereas with problems like plane safety we take an incremental and root-cause analysis approach to improve safety over time.

collision hot spots. It's hard to call them accidents when we deliberately build infrastructure that so reliably causes collisions.

AI learns to detect major intersections in satellite images

> satellite imagery of the area adds information about lane disposition, and the number of lanes, as well as the existence of a hard shoulder and the presence of pedestrians


A bit like autonomous driving, sounds like the sort of thing that probably works beautifully well in places like the US and elsewhere with wide roads and predictable grid patterns.

I imagine this sort of thing would mark large swathes of Europe,Africa, Japan, India and all sorts of other places as "accident hot-spots" though.

What I want to know is if it predicts bicycle accident hotspots -- especially angled intersections where at ordinary speeds for both vehicles, the angles are such that a cyclist can stay hidden from the drivers' view behind their right "A" pillar until the last second when it is too late. Accidents there are just a matter of odds, as whenever the timing is right, even good drivers will likely not see the cyclist until too late.

This happens in roundabouts where I live. The cyclist comes from the left in a curve and my car is also turning. With a specific timing and alignment of the movements the cyclist stays exactly behind the A pillar the whole time. Especially insidious is that by turning my car my view is sweeping over the roundabout but the cyclist still stays hidden.

I learned to swing my upper body sideways to have a look behind the pillar. As a cyclist I learned to look for the face of the driver and to stop if I don't see the eyes.

Because I both hit a cyclist and got hit as a cyclist by a car.

Yup, those road configurations can be really nasty - it's almost like setting up an illusion that becomes a trap.

Smart cars should improve this prediction by reporting hard braking and other evasive maneuvers that indicate "near misses" that wouldn't otherwise show up in accident reporting.

Which is probably good and bad -- municipalities can use the data to improve dangerous roads/intersections, online map providers could use the information to route around dangerous roads. But also insurance companies can use it to set rates "We've analyzed your driving over the past month and based on our predicted danger level, we're going to need to increase your rates 20%

I'm guessing you are against insurance companies charging more for drivers who are statistically likely to be more dangerous?

Car insurance telematics [1] seems largely positive in that it financially discourages dangerous behaviors and it means people who actually drive dangerously (as opposed to people who are simply young) will foot a greater fraction of insurance costs. (Though there are definitely implementation issues and privacy concerns.)

[1] https://www.consumerreports.org/car-insurance/how-car-insura...

There's a limit to how closely I want insurance companies (or any companies, really) to monitor me. Virtually sitting in the car with me is one of those limits.

While not related, I have always thought it would be interesting to plot police hot spots from Waze data..

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact