AI predicts accident hot-spots from satellite imagery and GPS data

https://www.unite.ai/ai-predicts-accident-hot-spots-from-satellite-imagery-and-gps-data/

by Hard_Space • 5 years ago

181 76 5 years ago

unite.ai

lixtra• 5 years ago

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

I hope they are misquoted here.

MauranKilom2• 5 years ago

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

batch12• 5 years ago

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

traceroute66• 5 years ago

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

Hmmm....

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.

notafraudster• 5 years ago

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.

simion314• 5 years ago

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.

theshadowknows• 5 years ago

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

mjmahone17• 5 years ago

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.

marcosdumay• 5 years ago

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.

pfisherman• 5 years ago

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.

h4kor• 5 years ago

AI learns to detect major intersections in satellite images

burlesona• 5 years ago

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.

Why?

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.

hetspookjee• 5 years ago

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.

mc323• 5 years ago

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.

ashtonkem• 5 years ago

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.

InitialBP• 5 years ago

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.

bryanrasmussen• 5 years ago

These are the futon of transportation.

Not getting the futon analogy?

boyd• 5 years ago

Futons are neither good couches nor good beds.

adrianN2• 5 years ago

So called Stroads: https://www.youtube.com/watch?v=ORzNZUeUHAM

goodpoint2• 5 years ago

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

lordnacho• 5 years ago

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.

throwaway0a5e6• 5 years ago

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.

airza• 5 years ago

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

what_is_orcas• 5 years ago

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).

prawn• 5 years ago

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.

flashfaffe2• 5 years ago

So is AI really needed here?

jvanderbot• 5 years ago

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.

brundolf• 5 years ago

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"

Mountain_Skies• 5 years ago

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.

MauranKilom2• 5 years ago

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):

https://ml8ygptwlcsq.i.optimole.com/fMKjlhs-dkVIYmPS/w:700/h...

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.

BoorishBears• 5 years ago

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

tbihl• 5 years ago

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

toss12• 5 years ago

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.

_nalply• 5 years ago

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.

toss12• 5 years ago

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

Johnny5552• 5 years ago

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%

asoneth• 5 years ago

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...

Johnny5552• 5 years ago

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.

AI predicts accident hot-spots from satellite imagery and GPS data

Related Stories