8 ms·

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

181 points by Hard_Space 5y ago

nikkinana 5y ago
Now how about border hopper hot spots?
traceroute66 5y 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.
lixtra 5y ago
> This is a step toward general AI, because our model can predict crash maps in uncharted territories.” I hope they are misquoted here.
- MauranKilom 5y ago
  The word "general" does indeed not appear in this sense in the paper.
batch12 5y ago
While not related, I have always thought it would be interesting to plot police hot spots from Waze data..
notafraudster 5y 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 5y 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 5y 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 5y 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 5y 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 5y 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 5y ago
AI learns to detect major intersections in satellite images
burlesona 5y 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 5y 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.
- bryanrasmussen 5y ago
  >These are the futon of transportation. Not getting the futon analogy?
  - boyd 5y ago
    Futons are neither good couches nor good beds.
MauranKilom 5y 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:556/q:auto/https://www.unite.ai/wp-content/uploads/2021/10/kernel-density-estimation.jpg 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 5y 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
QuarterReptile 5y ago
collision hot spots. It's hard to call them accidents when we deliberately build infrastructure that so reliably causes collisions.
toss1 5y 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 5y 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.
  - toss1 5y ago
    Yup, those road configurations can be really nasty - it's almost like setting up an illusion that becomes a trap.
Johnny555 5y 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 5y 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-insurance-telematics-discounts-really-work-a1549580662/ https://www.consumerreports.org/car-insurance/how-car-insura...
  - Johnny555 5y 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.

181 points by Hard_Space 5y ago

Now how about border hopper hot spots?

> 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.

lixtra 5y ago

> This is a step toward general AI, because our model can predict crash maps in uncharted territories.” I hope they are misquoted here.

MauranKilom 5y ago
The word "general" does indeed not appear in this sense in the paper.

batch12 5y ago

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

notafraudster 5y 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 5y 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 5y 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 5y 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 5y 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 5y 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 5y ago

AI learns to detect major intersections in satellite images

burlesona 5y 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 5y 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.
bryanrasmussen 5y ago
>These are the futon of transportation. Not getting the futon analogy?
- boyd 5y ago
  Futons are neither good couches nor good beds.

MauranKilom 5y 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:556/q:auto/https://www.unite.ai/wp-content/uploads/2021/10/kernel-density-estimation.jpg 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 5y 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

QuarterReptile 5y ago

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

toss1 5y 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 5y 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.
- toss1 5y ago
  Yup, those road configurations can be really nasty - it's almost like setting up an illusion that becomes a trap.

Johnny555 5y 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 5y 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-insurance-telematics-discounts-really-work-a1549580662/ https://www.consumerreports.org/car-insurance/how-car-insura...
- Johnny555 5y 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.