Maelochs

Oi.

Sixteen pages and we are on what, the third recycling?

Lady paid no attention. Sensors didn't work as well as they should have, or programming didn't pay attention to the sensors.

Unless there is some new info … go ahead ad recycle the same stuff again.

02Giant 

And yet, 480,000 people in the US die from smoking related causes, every year. Including 41,000 from second hand smoke.

Road Fan

You are mixing my response up with responses to a different question. I was responding to a discussion around Page 3 of the thread, related to "can AVs detect classic Driver Education scenarios, like groups of children playing near and on a sidestreet with parallel parked cars, when a ball rolls or flies into the road? As drivers we are expected and capable of intuiting when a kid is going to chase the ball. I have my doubts the AV algorithm can do the same, since it did not respond to the woman moving toward the path of the Uber. It should have responded because of its equipment with laser, radar, and optical sensors and a data fusion algorithm. One explanation for why it didn't could be that the person with her bike was not detected, but only a detailed analysis of the sensor data can tell us that.

That is the relationship between the woman who was killed and the children on a sidestreet. In both cases the system should have "reasoned" "there is a possible obstacle present and current speed and path is a collision course. Change the motion of the vehicle to eliminate the possibility of collision." The vehicle response in this situation should have been human-like: stop, slow down, steer out of a collision-possible path.

Since it had a monitor driver, it should have alerted the monitor driver to pay attention in the forward direction. Also because it had a monitor driver who was observed with an interior camera, it should have seen that the monitor driver's attention was not on the road (eyelids, eye gaze direction, not sampling the scene, and presumably no physical indication of engagement with path evaluation). Car companies have developed and deployed such internal observation cameras and driver engagement/drowsiness systems in the last 15 years - I know because I worked on them.

In such a test vehicle, if you don't support the monitor driver with sensing, monitoring and alerts, you risk that driver's natural but regrettable tendency for distraction allowing harm to occur to people who are not in the vehicle.

Road Fan

So @KD5NRH, she WAS an off-road object. But this is analogous to a vehicle-to-vehicle crossing-path collision where the vehicles should detect and track each other. This is the reason for full long-range 360 degree coverage by the scanning LIDAR. For some reason she was not indicated on the forward video until there was only a few seconds before impact so the brakes would not have had time to actually stop the car. But there should have been a function embedded called "Forward Vehicle Collision Mitigation" which is defined in an ISO standard, you should look it up. It says that if an object in-path is detected the brakes must be activated immediately even if the impact cannot be avoided. The ISO standard does not require off-path detection or anticipatory brake response beyond the core function, but not even this core was observed, seemingly.

If the AV is supposed to be at least as good as a human driver in a car equipped with a Forward Vehicle Collision Mitigation System, it has to have the capabilities required for that function with the robustness that will enable that function to be effective.

I-Like-To-Bike

The "forward vision" that you refer to is only the recording from the cheap crap dash cam that Uber installed on this vehicle for what ever reason. Presumably it had absolutely no connection to the software or hardware actually controlling the operation of the vehicle to include steering or braking.

I assume that the NHTSA investigators will look in great detail at this vehicle's hardware and software as well as the data gathered by the vehicle as well as that data transmitted to Uber, to determine what and when the vehicle detected the victim (or should have detected) and how it was programmed to respond to what it detected.

Road Fan

Agreed, but that's all that we mainly have access to at the moment. In my experience we added such cams to test cars with much more technical sensing systems to have a small set of summary evidence, at least that the test drive occurred and in general terms what happened.

So even though I suspect that pedestrian detection and path prediction did not happen, from this video there is no indication that the vehicle suddenly braked or that the steering wheel was suddenly turned to accomplish a path change.

At the same time, the monitor driver did show surprise, and we don't know if that was because she realized a crash was imminent or because the vehicle suddenly braked hard and lurched her. And we as the public may never know any of those things.

I-Like-To-Bike

If the Uber AV did not detect the pedestrian - Uber Fail!

If the Uber AV did detect the pedestrian, and still took no evasive/precautionary action due to its hardware/software inadequacies to operate safely on public roads - Uber Fail!

Road Fan

This is why I say the AV system needed to detect pre-collision conditions early enough to take action that is effective.

Maelochs

Which is why a a lot of hand-wringing about safety and health is ridiculous.

However ... in many cases (eleven out of twelve by your math) the people who die from smoking Chose to smoke.

To them, the reward justified the risk ... sort of like with drivers and cyclists.

Road Fan

Kind of straightforward, actually, but not simple.

Radar sends out radio waves that primarily are reflected by metals or electrically conductive objects. "Organic objects" have some electrical conductivity, but its low and variable. Radar is not usually good at seeing people, though it doesn't fail completely. Also the median looked very vegetated in the overhead photograph. Radar sometimes doesn't penetrate foliage, due to moisture and electrical conductivity, so she could have been effectively hidden from radar in the bushes. Same for the bicycle frame tubes and rims. So radar might not have been capable of seeing through the foliage.

Lidar is about objects reflecting visible light or infrared near-visible light that is safe to the eye. In general it cannot penetrate opaque surfaces. The bushes when leafy would have been pretty much opaque, and the LIDAR return signals would have been due to the leaves, not the person.

What I think should have worked perhaps even when she was in the foliage, is thermal imaging, far-inrared visual imaging. These can be very effective at isolating warm-blooded creatures, since they directly detect heat emanating from a living creature. No statement about whether the Uber car was so equipped.

When the person first emerged from the foliage, the far infrared (FIR) camera could have detected the person reliably, and lidar should also have done it if the person was in-range - she might not have been or her clothing might have absorbed the lidar energy and not reflected it.

So sensor effectiveness is not a given, rather it's pretty complex. Usually the resoution is sensor fusion, where the outputs of all these different sensors are combined, perhaps in a database, to make an object map that is more effective and reliable than any single sensor type. Again, it's not clear how the Uber car was equipped, but it's a pretty common solution.

McBTC

Let's not go too far and trot out bad science...


Secondhand smoke isn?t as bad as we thought.

I-Like-To-Bike

Or substituting WAGs (10,000 less fatalities! or any other fabricated statistic based apparently on tea leaf reading) about future safety benefits of driver-less cars based on speculation, wishful thinking, AV promoters' press releases and/or overheated rhetoric/bombast from anti-motorist zealots about the alleged ineptness/incompetence of human drivers in contrast to the enlightened, altruistic techno-wizards from Uber and Google and their ilk.

tyrion

You complain about "bad science" then link to a slate article? Are you serious?

McBTC

Is Slate a pro-second-hand smoke magazine?

Road Fan

Really, it should never be expected to be ok to run down a human being.

McBTC

Possible suicide– death by Uber?

Road Fan

But people without licenses can just call an Uber.

McBTC

What could you do if you were driving along within the speed limit and a j-walker dressed in black carrying a bike on a dark night suddenly stepped off a median into the path of your car? Is that something that you as a human with all that awesome brainpower at your disposal should obviously have anticipated?

tyrion

Slate is notorious for publishing bad science.

That "No Clear Link Between Passive Smoking and Lung Cancer" is the headline of a report of a single study. There are many studies that have come to different conclusions. And that study was about lung cancer and not about the more common threat of heart disease.

The consensus is that second hand smoke is harmful and causes death:

https://www.cancer.org/cancer/cancer...and-smoke.html

https://www.cdc.gov/tobacco/data_sta...acts/index.htm

https://www.cancer.gov/about-cancer/...oke-fact-sheet

McBTC

Interesting question although, not related to this case but more to human nature and all of the biased thought processes that people indulge in...

What if the pedestrian who was accidently run over and killed by an Uber driver had stolen the bike she was carrying at the time? Related?

McBTC

Slate does not believe science is self-correcting. Dr. John Ioannidis has pretty much shown science is broken --e.g. see, "Lies, Damned Lies, and Medical Science."

https://www.theatlantic.com/magazine...cience/308269/

CliffordK

Of course.

But, the Uber computer may well have identified the pedestrian/cyclist early and determined "not in my lane", and put it on ignore. Plus, of course, the car should have had the right-of-way. But, like you say, hitting someone for any reason is never appropriate.

A competent driver, of course, would have noted the pedestrian in the middle of the road (assuming the woman was visible), and watched her movement, estimated where she was heading, and slowed or moved the car to avoid the collision.

Hopefully the complete preliminary evaluation will be coming out shortly.

KD5NRH

Why wouldn't it also detect the ball and at least slow to reduce any issues from that roadway incursion? It's not going to have enough data to be sure it's a one pound, bouncy soccer ball and not a ~200 pound boulder at the front of a rockslide, after all.

Not likely; ever used NIR night vision? Plenty of clothing that's black to visible light is highly reflective to NIR. The range on the LiDAR unit is 120m, so even if it failed to predict her path, she was in the AV's lane well within LiDAR range, with no close background to mask her return. Even if it didn't know what she was, why wouldn't the algorithm try to avoid any solid object in the roadway whenever possible? It wasn't being tailgated, so no undue risk in maximum braking, and the other lanes were empty, so swerving was definitely an option as well. Even going over the curb and potentially into a ditch or small tree should be a far preferable alternative to hitting something roughly 5' tall and about the same wide that just might be alive.

tyrion

Even if one agrees with Dr. John Ioannidis, that in no way supports cherry picking studies to support a desired conclusion.

I-Like-To-Bike

Who needs studies at all when WAGs and specious arguments from the AV promoters are available in press releases to "support" predictions of significantly reduced injuries, damage and fatalities?