Fusion of Stereo Vision for Pedestrian Recognition using CNNs [pdf]
hal.inria.fr
hal.inria.fr
Approach: Train a CNN-based pedestrian classifier using a combination of intensity, depth and flow modalities on the Daimler stereo vision data set. The resulting CNN classifies objects in images as "Pedestrian" or "Non-Pedestrian" (rather than the more granular goal above).
Result: Outcomes are measured by the False Positive Rate (FPR) at the 90% True Positive, or detection rate. Lower FPR=better.
The CNN created in this experiment ("Late Fusion") achieves an FPR of 0.0125%, which is almost 50% better than the 2010 "Early Fusion" CNN's 0.02% FPR over the same dataset.
Still, a 2011 model ("HOG+LBP/MLP, Monolithic HOG classifier") that uses "explicit features" vastly outperforms both, achieving an FPR of 0.00026% over the same dataset.
Conclusion: "We showed that the late-fusion classifier outperforms the early-fusion one but does not go beyond the state-of-the-art."
On a side note, I am reminded of an anime called Psycho Pass [4] which uses psychological rating given by a super intelligence to decide who needs to be taken off the streets to be treated.
-------------------------
[0] https://www.newscientist.com/article/mg21528835-600-cameras-...
[1] https://en.wikipedia.org/wiki/Gait_analysis
[2] https://en.wikipedia.org/wiki/Sentiment_analysis
[3] http://www.popsci.com/byzantine-science-deceiving-artificial...
Imagine the airport of the future where they won't bother with "Mr Smith, please report to Gate 6 for final boarding" while Mr Smith sits oblivious at the bar. Instead it will be "Mr Smith? We brought you your bill early because your flight is leaving in 10 minutes and you should get going" "But I didn't tell you my name or flight..."
"Would you like another drink sir?"
Why alert him when they can sell his seat, sell him a few more drinks, a massage, and some entertainment?
This is a legitimate reason to worry about AIs. The "world domination" angle seems overblown, but machine learning has the potential to evolve some immoral behaviors like this without anybody even knowing.
I imagine this sort of transaction going on a lot in the future. It kinda happens now via email and in browsers. My browser sort of organises and filters all sorts of messages from websites. Also some primitive geo-location also happens... this can only get better as the services mature.
When are we going to start seeing methods for measuring intent or theory around guessing the future nature of an object based on contextual understanding of that object? How does a car recognizing a pedestrian help if the pedestrian starts running toward the car?
It just feels like there's a lot of patting of backs that happens around this stuff, when we really aren't even close until we have a system that has as learning and understanding approach that is as abstract as a human's.
Our current roads are OK for human drivers, as they are designed heavily around visual signals, we can process and synthesize a lot of visual/audio/learned data well.
Machines can process visual data faster than us, but can't synthesize it with other data nearly as well as we can. Adding sensors that passively communicate information, such as the location of lanes, other cars, and humans, will reduce a lot of the issues and edge cases of visual-based ADAS systems.
I could imagine a future where cars are always listening for smartphones (or other wearable/implanted sensors) in "pedestrian" mode, which could be used to override regular function and stop the vehicle if detected within a certain distance, or intersecting at a certain trajectory.
[1] http://karpathy.github.io/2012/10/22/state-of-computer-visio...
Leaving the USA will actually ensure the NSA collects a lot more about you. Staying will actually ensure they collect very little, mainly as a byproduct of the methods they use and not because of any nefarious plot.
Even if you manage to get your government to never do this, it is (or soon will be) easily within reach for private hobbyists.
My window faces a major freeway. I could probably rig up a license plate reader that registers all cars going by fairly cheaply. If not now, then surely in 5 years.
Everyone has a quite nice camera in their pocket. The car I bought last year comes with 3 cameras. And so on. I think it's an unavoidable fact that in the very near future, you have to assume you are being filmed whenever you're in public.
So I think the real question is: How do we adapt to this new world? How can we limit the bad aspects and enjoy the good aspects? Are we really sure what is bad and good?
1) It's a lot easier to recognize 100M distinct plates than faces. I don't see anyone having the resources and patience to do the latter any time soon. Even the FBI isn't claiming they can recognize 1 face/100M with low error today. At best, they have such a database of pictures. I'm have big doubts that usefully matching against so many faces is even technically possible.
2) I think you're right, given the rapid advancement of tech, our emphasis shouldn't be whether we should outlaw surveillance practices and methods. Instead we should focus on controlling access to such databases. Who gets to gather or use such data and when should it be authorized? Should we require that each ID match request be explicitly granted (like a warrant) or should we at least require that all such requests be logged, and perhaps publicly reported?
Once access to sensitive data is regulated, it no longer matters how the data is gathered. Abuse of such a law should poison any fruits of the data, at least on the part of lawful agencies.
For example, in WA State, you can cross at any unmarked crosswalk which is defined as any intersection (plus a few non-intersections), unless there is a sign explicitly disallowing crossing. I wouldn't know how to accurately estimate this, but there are probably over 100k unmarked crosswalks in Seattle alone (most of which are ignored completely by humans). The pedestrian's responsibility is to not enter the roadway without enough time for a car to stop. The car's responsibility is to stop once the pedestrian signals intent to cross. Signaling intent does not require waiting at the edge of the roadway...merely walking towards the roadway while on the sidewalk is considered signaling intent to cross.
What this effectively means is that the car must do far more than understand pedestrian behavior in crosswalks. It must understand the periphery of the road...not just the pedestrians it can see, but the pedestrians it can't. Most speed limits are actually too high to take into account legal pedestrian behavior. The speed limit is merely an upper limit; driving too fast for conditions can always override the speed limit as a factor in an accident. This means that trees, buildings, etc., that obscure view of and behavior of pedestrians on the sidewalk implicitly lowers the speed at which cars can drive on that road.
Now today we essentially get away with favoring the driver. The driver can assume that they were acting reasonably by driving the speed limit, and we can say that the pedestrian was out of line by walking too fast out into the street, and most situations are resolved blindly based on testimony with no evidence. But the risk model of autonomous cars is different. The manufacturer is gonna be on the hook, not the driver, and the car needs to be able to obey the letter of the law with a higher burden of proof and with dozens of sensors recording the situation.
As we can see from this paper, pedestrian recognition is still a hard problem. It's hard in the engineering sense, and the mathematical sense. But cars don't need to just recognize pedestrians...they need to recognize and understand the context in which a currently invisible pedestrian could appear in the near future. And perhaps more importantly, they still need to solve the problem of instantaneous path planning in a 2d space with a dynamically changing 3d obstruction model consisting of far more than just pedestrians. It's easy to look at the past and say we've made rapid progress in the past so we'll make rapid progress in the future...but in the process of maturing a technology we regularly see exponential progress in the beginning and asymptotic progress near the end. This is going to be very difficult.
I imagine most attacks aren't people being like "You know what, I'm sick of life. I'm just going to ram my truck into people"
https://www.youtube.com/watch?v=HnOPD2MOngU Here is a small engine whose timing is controlled by an Arduino. You vastly overestimate what can be done, and underestimate the intelligence and willpower of terrorists.
All you're achieving is making it harder for everyday people to work on their vehicles.
If you want to make every aspect of life feel like commercial air travel does at present, then sure, then play your ever-escalating game of whack-a-mole, but please don't speak as if the conclusion is so foregone. And don't pretend the benefits will naturally materialize (when history has shown that they won't). The cost of such a system is the only part of it that's guaranteed.
"The trouble with you Americans is you think every problem has a solution," it has been observed.
I'm sure this is 100% doable right now with existing technology.
I also could think of anti-human-targeting weapons to stop or neutralize those vehicles.
Everything is possible.