Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

This.

I'm always surprised how few people know as pedestrians to look at me, the driver and not my car. Some don't even look at the car. I developed a habit of looking for their gaze and if they don't look back, assume they're not fully aware and just am more cautious.

This works because I, as a human, know this and can compensate when they just rush the crosswalk without being fully aware of their surroundings.

How do you do that with a machine?



Just because a pedestrian is looking in your direction, it doesn't mean they can actually meet your gaze. I don't drive and always try to make eye contact with drivers, but tinted windows and windshield glare often make it impossible in reality.

So I will stare at where I know the driver's face ought to be, but I can't actually tell whether they have seen me. Tinted windows darken the inside of the car and that makes windshield glare all I can actually see.


It's probably equally simple for a machine (if not simpler) to figure out if a human has looked at the oncoming vehicle with sufficiently advanced cameras and computer vision. However, we use a lot more hints (change of pace, facial expressions that indicate presence and focus, nodding etc), which, while not outside the realm of AI CV, would surely need lots of training, yet it comes pretty naturally to humans.


The problem is not whether the machine can see the human, or even whether the machine knows that the human has seen the machine.

The problem is that the human currently has no way to know whether the machine has seen the human.


The GP was talking about the driver knowing if the "human has seen the machine".

For the inverse problem, we could simply start adding screens (instead of windshields?) to self-driving cars that acknowledge the pedestrians in a particular way (when there's only a few people, in the Black Mirror realm, they'd actually greet them by name using facial recognition and universal DB of everyone :).


What? I was absolutely talking about human cyclists and pedestrians needing to know that the machine sees the human, and not by blind faith, but through some active explicit demonstrable indication.


I believe my reply was to user LoLFactor (threads open up subthreads which open up more subthreads...).

If you are disagreeing with my use of "GP" (grandparent post), always go with the HN rule of "assume the best possible interpretation" :)


Of course there is, the behavior of the machine should be the same as of a driver - the car shaped object starts to slow down in a way that will make it stop before hitting me at the pedestrian crossing.


This one scenario is an entirely insufficient list of all possible situations that occur.

Some of the worst incidents are when a stopped car stars moving just for the most obvious example.


It is not possible for me to see you in your car with any sort of reliability; even if I could the benefits are dubious and again, unreliable.

Looking at your car is all I need anyway - I can tell if you've seen me by your behavior, you're either slowing down to yield to me or you're not. If you're not, the only possible outcome of knowing you're seeing me is being misled into stepping into your path of travel.


It's not all or nothing. A hundred different things all add up and none of them do the entire job, nor are entirely unnecessary simply because they don't do the entire job.


Hey I'm always surprised at how many pedestrians walk around all in black (or dark) with no light, no reflecting stripe, no nothing and no care in the world. Even with the best of intention of drivers, it's a death wish. And it's everywhere.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: