EEVblog #1068 - Autonomous Uber Incident Update

[james_s]

Having ridden in a convention taxi cab a few times, it did not exactly inspire confidence in the driver training and abilities, I'd just as soon hop in a car with a random stranger.

[Clear as mud]

I just saw this this morning.  I'll probably come back later and read the rest of the thread, but I noticed there were a couple of posts on the first two pages saying they could not figure out where the bicycle came from at 6:17.  It can be seen in front of the van parked on the left at 6:17.

I went back and watched the video a few times and figured out where the cyclist came from.  He can be seen at 6:13 heading towards us, just beyond the second crosswalk.  He only appears for a second or so between pedestrians and the automatic system does not detect him at that time either, but he's still pretty far away.  After that he is eclipsed for about three seconds by the man walking diagonally from right to left across the road.  The cyclist is approaching from further down the whole time, but most of the time he's behind that other pedestrian, then after passing the two parked vans, the cyclist turns and goes toward the sidewalk, at which time he comes out from behind the pedestrian and can be seen again (6:17 in the video).

[apis]

I just saw this this morning.  I'll probably come back later and read the rest of the thread, but I noticed there were a couple of posts on the first two pages saying they could not figure out where the bicycle came from at 6:17.  It can be seen in front of the van parked on the left at 6:17.

Indeed, I had to go back an watch it again to see it. It's interesting since it's a good illustration of how we humans also fail at this task. And we could watch it from the comfort of our chairs while sipping coffee and pausing and going back and forth in the video. The algorithm has to run in real time.

[Brumby]

How can you say we humans "also fail at this task"?

Perhaps I should get clarification on exactly what the task is.

IF it is the identification of every object within one's environment, then I would challenge that task as not even being reasonable.

As far as performing a risk assessment is concerned - a much more pragmatic task - then that bicyclist would have been ignored by most people - as they never really entered the scene as any real risk.

Let's not hold the bar so high that we get lost in detail that holds little significance in the whole picture.

[apis]

I completely agree. We should look at the big picture instead of demanding perfection from every detail.

IF it is the identification of every object within one's environment, then I would challenge that task as not even being reasonable.

Indeed, my point was that demanding that this collision avoidance system can identify every object is unreasonable and that humans can not do that either. What matters is that it can reduce the number of accidents (which it apparently can, significantly so).

A robocar must be much more reliable than a collision detection system (which is why they do not solely rely on this kind of computer vision system). But it is also unreasonable to demand that robocars should be perfect before admitting they are useful. What the manufacturers need to prove is that the cars are safer than human drivers, not that they are infallible (which they of course never will be).