A person on a skateboard is still a person, would any AI trained to spot people not recognise that (at the same level as it spots people in general)?
Do you really need to train your AI to spot people on skateboards vs rollerskates vs rollerblades vs heelies vs sliding on ice vs traveling on a travelator vs standing on a moving vehicle vs ...?
OK, being able to recognise the differences and act accordingly might be useful but the principle of "person getting closer to vehicle" should hold sway for most situations?!?
The skateboard is relevant because it could theoretically make a neural network fail to identify a person that needs to be avoided
>human moving towards you at X speed
Lidar systems will work this way. Camera+AI not necessarily, it will still need a way to sense relative speed otherwise you are banking on an AI to identify an obstacle in an image.
To put it in more direct terms, I wnder if AI training does and/or can even distinguish between different classes of pedestrians. I am suspecting the current approach is to treat them all the same. Which would result in a driver that doesnt know and therefore violates the "Vertrauensgrundsatz", someone who wouldnt even get a drivers license where I dwell...
In case of a car, the thrown object is the easy case.
Now imagine you spot a person walking down the road on a pavement. Suddenly, the person turns towards the edge. What do you do?
See, a human driver would look at the gait of the person from far away to evaluate for instance if they're sober, or if it's a child who might be expected to run in. Whether it is near a crossing or a potential crossing. Whether the person was walking or standing... Many other obvious and less obvious indicators.
AI currently sees a moving blob of pixels in a shape of a person. No advanced inference.
When interviewed in case of an accident, say because they got rear ended due to braking, a person can explain why they braked most of the time.
AI now couldn't even say which features it weighted.
Yes, except this presumes that AI would recognize the kid at the same time as a human driver. Or at least soon enough such that AI's quicker reflexes can compensate for any delay.
If AI can't recognize some obstacles at all, then I think it's safe to assume there may be some delay in obstacle recognition compared to humans.
Yes. My point is that detecting people is not hard and everyone but Tesla already seems very good at it. What's hard is understanding what they want to do and reacting to it. That part is much easier if everything is going 10 mph or less.
Detecting inconspicuous transition between sidewalk and road is difficult, but if you haven't noticed, human drivers don't seem to care much about it either.
There's a whole lot of challenging information that is completely natural and intuitive for a human to understand but fiendishly difficult for a ML algorithm to figure out. There's some cues that I'm sure we probably won't be able to use until we create a genuine artificial general intelligence. If you're driving and you see someone standing at a crosswalk it's intuitive just by looking at them if they're waiting for you to pass, about to walk out into the road, panhandling, etc. You can put yourself in their shoes and make a reasonable prediction as to what they are thinking about doing. On a previous HN thread there was a commenter that had a run in with a Waymo car while riding a bike. He was coming up to a 4 way stop, and yielding to the Waymo car. If he was balancing on his pedals, still stationary, the Waymo car would stop and not proceed through the intersection, apparently interpreting his pose as if he's about to proceed through the intersection itself. A human driver wouldn't have an issue with that, but you can imagine the training data would show that when a bicycle is stopped, the rider puts a foot on the ground.
Generous helpings of lidar and radar to augment cameras is a crutch to help compensate for the lack of 500 million years of unsupervised learning that went into our visual cortex.
A machine can already do the act of stopping better than 100% of the people.
The problem is recognizing the obstacle with a better accuracy than people do. Tesla's a good example for this, and being unable to recognize certain obstacles in certain conditions.
Not in the field but I am assuming that the AI/ML cannot handle the ability to see possible road hazards with in the human peripheral vision area
Such as a kid attempting to chase down a ball that is rolling towards the road (object vector path collision), specially with the ball and or kid suddenly hidden from view because of a parked car (real environment vs visual environment). Or a group playing basketball in a driveway. Both where slowing down is always the safer bet.
The car would have to solve problems like: "are those people on the side of the road behaving in a way that would suggest they might accidentally fall into the road". You or I might realize that a kid trying to treat the curb as a balance beam is entirely different than one walking in the middle of the sidewalk, but to do so takes some understanding that I'm not sure AI can learn by watching lots of driving scenarios.
The problem is that something as ambiguous as you described requires human-level AI.
But if you could accept portions of it, you could build a detector for unusual vehicles (would yield a lot of false alarms). And a detector for beautiful sunsets. A system that would recognize individuals by gait and be trained by you for specific authorized persons. A detector for water on the ground.
It's possible to just detect any movement or significant change in the scene, but that would yield quite a few events that were actually not interesting. Like birds landing.
It's great that this can recognize pedestrians, but I've still seen no theory around decision-making. This last week, I happened upon a tree that fell into the road after a bad wind storm. I _know_ a tree when I see it, and I know it is large and heavy and difficult to move and would take a great deal of time for that to happen, so I immediately cut to a route around it. But what if it was a bus or truck turning around? Well I know that a bus turning around is large and heavy and difficult to move, but it will be moving out of the way in a moment, so I can wait for that to happen.
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.
Yesterday I was on a two lane country road where I saw a man standing on the opposite side of the road holding up a bright orange baseball cap in his hand and his other hand in a "stop" motion. I slowed down enough to where I didn't hit the crashed car which was just around the bend, but I may have hit it without the advance notice because the turn had very poor visibility. If I were an AI driver, I probably would have considered the man "not-an-obstacle" and promptly ignored him and hit the other car. After all, I'm trained to slow down when I see people in hi-vis with proper "slow traffic" signs. Some random person whose neighbor crashed and is waving the brightest color thing they have is not in the training data.
Or in other words, there are signals humans can send to other humans which can be understood even if you haven't seen that particular signal before, and those will be the biggest challenge for an AI driver, at least if it resembles the current generation of AI.
But you, as a human, can look at another human walking parallel to your car and based on your life experience and knowledge determine the probability of that human ending up in front of your moving car. And you'll do this in microseconds. And you can do this for every other object in your vicinity.
Especially if it has AI and recognition systems powerful enough to make these decisions reliably in the first place.
We're imagining a scenario where a computer is capable of reliably recognizing that someone is an elderly female doctor, and is capable of responding to that information in milliseconds before slamming into that person, but the car is somehow not capable of recognizing ahead of time that a human is on the side of the road and about to step into it.
Cars that have AI systems advanced enough to recognize what profession a pedestrian is in should also be able to recognize that pedestrians are on the side of the road, or that the car is entering a blind spot where a pedestrian might enter the road. These are very selectively omniscient cars that the website imagines we've built.
There's no need for a theory of mind - the self driving car can identify the kid as a pedestrian, recognize that it started moving in a possible collision course, then disappeared, thus prompting either slowing down to a non-fatal speed until the truck has been passed.
Children playing in the street are a common occurrence in residential areas, I see no reason why you would not develop a set of rules and heuristics to handle them. Identifying a pedestrian as child, and knowing whether it is running or playing, is well within the capabilities of modern computer vision.
Do you really need to train your AI to spot people on skateboards vs rollerskates vs rollerblades vs heelies vs sliding on ice vs traveling on a travelator vs standing on a moving vehicle vs ...?
OK, being able to recognise the differences and act accordingly might be useful but the principle of "person getting closer to vehicle" should hold sway for most situations?!?
reply