Google Research Leads To Automated Real-Time Pedestrian Detection

An anonymous reader writes with a link to a story about one of the unexciting but vital bits of technology that will need to be even further developed as autonomous cars' presence grows: making sure that those cars don't hit people. Google researchers have recently presented findings about a method that tops previous ones for real-time pedestrian detection using neural nets "that is both extremely fast and extremely accurate." From the article: There are other approaches that provide a real-time solution on the GPU but in doing so, have not achieved accuracy targets (in this real-time approach there was a miss rate of 42% on the Caltech pedestrian detection benchmark). Another approach called the VeryFast method can run at 100 frames per second (compared to the Google team's 15) but the miss rate is even greater. Others that emphasize accuracy, even with GPU acceleration, are up to 195 times slower.

Comments

Re:So...

[paskie]

No. As usual, the summary is confusing as it gives numbers for the *older* methods, but the current Google's method is: "The resulting approach achieves a 26.2% average miss rate on the Caltech Pedestrian detection benchmark, which is competitive with the very best reported results. "

So, there's 26.2% chance that on a single particular image, you miss the pedestrian (at the same time, it seems that in about 15-20% images it sees a pedestrian that is in fact a shrubbery or whatever). This is an academic dataset, and in reality you will have a video feed. AFAICS it's not clear how the precision translates when you have a sequence of many pictures of the pedestrian - whether you will have much higher chances to spot them at least on some of them, or if it's more of a systematic problem and khaki-clothed people just don't stand a chance.

Two ideas

[CustomSolvers2]

Idea 1: the fact that accounting for the eventuality of not hitting pedestrians (or any other being/thing) is "one of the unexciting but vital bits of technology" when talking about an autonomous car provides a quite accurate summary about what a big proportion of AI-focused approaches are about. Lots buzz (= exciting technological break-troughs) and not actually-working results (= unexciting technical bits avoiding the big idea to work at all). And this is not just what the OP thinks; Google has been testing autonomous cars for some years already without having still tackled such a secondary(?!) issue.

Idea 2: after quickly skimming through this paper, it seems that the new much-more-accurate algorithm still misses 30% of cases. For me, hurting (even killing) 3 out of 10 pedestrians still sounds quite bad. Additionally, we are talking about their training dataset whose exact complexity is not too clear. For example: what about a kid suddenly crossing the street?, how good is this algorithm at differentiating between persons and similar shapes (human-like advertisement)?, how does it behave in poor-visibility conditions?, etc.

I am completely aware about the tremendous difficulty associated with accomplishing the expected goal and the outputs so far seem promising. But why are they implying that something is almost done, when quite a few basic problems haven't still been tackled?

Re:Two ideas

[Gravis Zero]

it seems that the new much-more-accurate algorithm still misses 30% of cases. For me, hurting (even killing) 3 out of 10 pedestrians still sounds quite bad.

missing 30% isn't killing 3 out of 10 people, it's killing 7 out of 10 people which is a solid 70 points or 210 points if you are drifting. #Carmageddon

Re:Two ideas

[WoLpH]

Well, the question here is. Given that 26.2% at 15 fps, does that make the probability of a detection within a second (1-.262)^15 = .010491689? So less than one percent?

Or is it far larger because the results are not actually independent?

Re:Two ideas

[CustomSolvers2]

I took the information from the paper itself, more specifically from the following part:
"For example, when training on the KITTI pedestrian dataset [18], the best known average miss rate is 61.2%, whereas when training on INRIA [10], the average miss rate is 50.2% [6]. Both miss rates are much higher than 31.1% of our method"

As I understand it, a miss rate of 31.1% on a given dataset means that 31.1% of the tested attempts failed.

Re:Two ideas

[umafuckit]

For me, hurting (even killing) 3 out of 10 pedestrians still sounds quite bad.

Unless we know what the video feed is we can't make that statement. Are these pedestrians crossing the road or on the sidewalk? If the algorithm is missing 3 out of 10 sidewalk pedestrains that's much less serious than 3 out of 10 crossing the road. I suspect the idea behind the visual search is to identify people who could potentially cross the road so the car can slow down in anticipation. People actually on the road, in front of the car, can be spotted in other ways using other sensors.

Re:Two ideas

[Anonymous Coward]

But why are they implying that something is almost done, when quite a few basic problems haven't still been tackled?

This isn't about not hitting pedestrians in the roadway. This is about categorizing objects outside the roadway so that you know if they are pedestrians, who may enter the roadway at any time, or stationary objects, which may be presumed to stay stationary. Once categorized as a pedestrian, additional algorithms can be used to guess whether the pedestrians are going to enter the roadway and take defensive driving precautions to avoid hitting them.

Note that even if this algorithm were wrong 100% of the time, the actual injury rate would be much lower. While pedestrians can enter the roadway at any time, most stick to the sidewalk most of the time. Even those who enter the roadway will generally be trying to avoid being hit by cars. So normal obstacle detection (which doesn't care if they are a pedestrian or an inanimate object) will be able to avoid them most of the time. Which may help explain why Google cars haven't been hitting 26% of the pedestrians they pass but instead 0%.

In other words, this isn't quite as basic of a problem as you make it out to be. This will make no difference in the vast majority of trips. In a small number of trips, it may cause the car to take extra precautions to avoid potential interactions with a pedestrian. Even there, most of the time it won't matter (i.e. the vehicle would not hit the pedestrian even without the precautions). In a very, very small number of trips, this could allow an autonomous car to avoid hitting a pedestrian that a human driver would have hit. In a truly miniscule number of trips, this could allow the autonomous vehicle to avoid hitting a pedestrian that a human driver would have avoided but the autonomous vehicle otherwise would not have.

Even without this, autonomous cars are likely to be safer than human drivers. The biggest contributor to accidents by human drivers is negligence. The driver makes a mistake, is distracted, or whatever. Those things don't happen to AI. I'm not saying that we shouldn't tweak the AI to avoid a one-in-a-million accident. We should. But it's not an absolute bar to deployment. Presumably we will continue to tweak and improve the AI after deployment.

Actually, the biggest problem that I see with this is that they are using neural nets to "learn" the difference between pedestrians and other objects. It's much harder to fix a neural net problem than an algorithmic problem. You can tweak an algorithm. With a neural net, changes are as likely to make the situation worse as better. This means that we don't fix the system any more than we would fix a human driver who was in an accident. Neural nets are just too complicated to tweak -- that's why we're using a neural net rather than an algorithmic solution in the first place.

It doesn't have to be perfect.

[SuricouRaven]

It just has to be equal to a human driver - and human drivers are not that good.

Re:It doesn't have to be perfect.

[pellik]

No, it has to be way better then a human driver. There is a completely different scale of liability for a self driving car then for a human. The blame will hit much harder when someone gets hurt.

"Active City Stop"

[Kartu]

One of the recent models of Mazda I drove (I'm pretty sure all manufacturers have that, Ford at the very least) had "active city stop" feature, active at speeds up to, 30km/h, if I remember correctly.

Car would emergency break ON ITS OWN if it would spot a pedestrian.
To my knowledge, they use some "radar like" technology for it.
I guess it's not far sighted enough for a self driving car.