Viewing 26 reply threads
  • Author
    Posts
    • #2543

      damon
      Participant

      The video is shocking…

      https://www.theguardian.com/technology/2018/mar/22/video-released-of-uber-self-driving-crash-that-killed-woman-in-arizona

      Before seeing this, I’d given the self drive car the “benefit of the doubt”, kind of imagining that the woman had stepped of the curb a nanosecond before the car passed, meaning the accident was totally unavoidable, regardless of the technology. But it’s not the case. So for me it kind of answers a few questions…and raises a few more. “Could a human driver have avoided the woman”? I’d say after seeing the video the answer is “no”, because she was walking across an unlit road, at night, with a bike with no lights, wearing dark clothes… But it raises the question: “Shouldn’t we expect a self-drive car to perform better than a human precisely in this type of situation?” I would have imagined that the technology they are installing provides the advantage that self drive cars are able to “see” dangers at night that humans can’t see, which was clearly not the case.

      Still, we ain’t gonna be seeing any news reports of situations in which self-drive cars have saved lives, of the potentially overwhelming number of “non-accidents” that are going to be occurring thanks to improved safety…

    • #2544

      Steve
      Participant

      Tragic case. A couple of things spring to mind, firstly I thought autonomous vehicles used laser scanning or some other similar system, the fact that it was dark should have been totally irrelevant. Secondly, although it’s difficult to tell from the video, there appears to be no attempt by the vehicle to brake or try and avoid the collision. There must have been a serious failure somewhere.

      The interior camera is also interesting. I always thought the idea that at this stage a car would be autonomous but the driver would be able to ‘take over’ was nonsense. It’s obvious that once the idea that the car is on autopilot becomes accepted by a driver they’ll be daydreaming, snoozing and picking their nose with zero idea of what’s happening, and this is clearly the case.

      Clearly the technology is impressive, but society will have to accept there will always be accidents, even when the technology is perfected. Autonomous vehicles will be safer than humans, but we shouldn’t fetishise technology as being completely infallible, which some proponents seem to belive is the case.

    • #2545

      sammo
      Participant

      I think the guy in the car is going to get a lot of fallout over this but its an unreal expectation that anyone could stay 100% focused. Im sure that the guy has been in the car for tens of thousands of accident free miles and that breeds complacency.

      Maybe testers should be limited to say 1 hour stretches, there is Mobile phone tech that keeps the screen on if the camera detects your eyes on it, maybe the self drive cars during testing should have something similar.

      I can and often do drive for 8 to 10 hours in a day with no problem but who could sit for that time not doing anything but looking forward?

      On to the car, I too thought they had some kind of lidar / radar / infrared / night vision etc so something must have failed, possibly multiple somethings for that to happen but on balance from that video the fault falls on the pedestrian. It was dark, the car had lights on and she walked out in front of it.

    • #2546

      damon
      Participant

      @steve Every new innovation throws up a whole host of new risks, and we just grow to accept these new risks as the price to pay for the enormous convenience the innovation offers: Flight, cars, the internet etc…

      But it’s going to be very interesting to see how the new issue of risks presented by autonomous cars plays out. It’s different in the sense that it’s a new management of an existing technology & people seem ready to demonise it as “bad” despite the probability that it will improve road safety enormously (IMO). And also the fact that the fault will shift from driver error to manufacturer error…so perhaps we’ll see car insurance come down enormously, while insurance policies for the car companies will skyrocket…?

    • #2548

      Jethro
      Participant

      @damon I seem to recall Audi having human/animal detecting sensors as an option some time ago. Their adverts made reference to it. I dont know if the braking systems were integrated with it though.

    • #2549

      Jamesg85
      Participant

      @damon I think a human (and computer) should have seen her in the headlights 50m ahead. Cameras don’t record what we see outside of the bright areas and unlike real life the internet video is not 3d and high resolution.

      You really should be able to avoid a slow moving human on a clear road.

      *what’s the range of headlights? 100m or more to see something human sized otherwise we’d be crashing into postboxes and dustbins every night?

    • #2550

      daveb
      Participant

      Wonder what the crash stats are per 1000 miles driven etc… people will die in car accidents in every country today, some will have already; they just don’t make the news.

    • #2551

      dantheman
      Participant

      At the risk of attracting a flurry of anger….

      Could a human driver have avoided the woman

      Could the woman crossing the road have avoided the traffic, driverless or otherwise, by looking before crossing the road?

      I think a human (and computer) should have seen her in the headlights 50m ahead.

      And the woman crossing the road should have seen the headlights much further away than 50m.

      Secondly, although it’s difficult to tell from the video, there appears to be no attempt by the vehicle to brake or try and avoid the collision.

      There appears to be no attempt by the pedestrian to look or try and avoid the collision.
      Roads have vehicles on them. If you’re going to wander across them without looking, something will hit you eventually, regardless of what or who is driving.

    • #2552

      Jamesg85
      Participant

      @dantheman On the video it’s sudden but in the real world is not low quality video – you see pavement, kerbs, other lanes, potholes and 100m+ on an open road at night or shape of the road much further. Otherwise you’d crash at every bend unless you crawled along.

      Roads have pedestrians, vehicles, potholes etc on them. If you’re going to drivre along them without looking, you will hit something eventually, regardless of what or who it is.

    • #2553

      Ricky m
      Participant

      I am very surprised the sensor suite didn’t pick the pedestrian and bike up in time.

      Hopefully all the sensor data is available so that this can be learnt from in a measurable way that is pushed out to the whole Uber fleet and disseminated to other manufacturers by the relevant federal agency (NHSTA or some such), which at least is a much more usefil outcome than with a human driver where no lessons are learnt by the wider fleet of human drivers.

      Perhaps it’s just the interior camera angle, but the safety driver seemed to spent a lot of time not looking at the road. If so, this raises serious questions about their training and Uber’s continuous monitoring of their engagement. Given Uber’s general business ethos (I can’t bring myself to say ethics) this sadly doesn’t surprise me.

    • #2554

      Bear
      Participant

      I wonder how good the Lidar is on the Uber cars is (or if they have Lidar or are trying to get by with radar and cameras). Uber seem to have more accidents than the other companies running test programs and there is a lawsuit between Uber and Waymo (Google’s self driving car company) about Uber hiring Waymo engineers to steal Lidar technology.

      At this point in the technology development I don’t think we can assume one self driving car is pretty much as good as another. It’s more likely that some companies are significantly more advanced than others and also some will be aiming for lower price points and be trying to get by with less expensive sensors and computers. Maybe the Uber image processing system had difficulty recognizing an object which was a combination of person and bicycle in that context.

      We also shouldn’t just assume the driver was dozing off. He may be looking down at a monitor showing data from the self driving system or the car’s view of the road ahead.

    • #2559

      bob
      Participant

      That’s awful, I’m surprised it’s been released. It’s shocking, not because it’s a computer in control nor because the (a) human is obviously incapable for totally forseeable reasons of reliably undertaking a back-up safety driver role but because that accident could happen to any one of us, thankfully for most when it does it’ll be a deer and a car not a woman we kill.

      I’m not familiar with UBER’s test vehicles but I’d be surprised if there wasn’t multi-spectral imaging, at least visible light cameras and radar/lidar. It’ll be technically interesting to understand how and why it failed. Until that’s understood and hopefully (though perhaps not necessarily) rectified it’s reasonable that the vehicles are withdrawn from live tests.

    • #2560

      Keith
      Participant

      I am very surprised the sensor suite didn’t pick the pedestrian and bike up in time.

      I’m not. A fair bit of information has been published about the difficulty that autonomous vehicles have in detecting cyclists (Google “autonomous vehicles poor at identifying cyclists”). A person pushing a bike should have a similar radar image to a cyclist, I’d have thought – better, possibly, given that the victim in this case was travelling perpendicular to the direction of travel of the vehicle so the full face of each wheel was presented to the sensors. But it seems that the car still didn’t ‘see’ anything – or if it did, it didn’t react in any meaningful or useful way.

      There is a view that what might at present appear to be an inherent weakness in the technology could lead to vehicular roads being ruled off-limits to anything smaller and slower than a motorcycle. Interesting to speculate what that could lead to in terms of the street environment in your average city or town centre (in the UK, anyway – in the US pedestrians already seem to be widely treated as pariahs anywhere other than in motor-vehicle-free zones).

    • #2581

      eric4
      Participant

      I spent a couple of nights in Flagstaff Arizona a while ago. The street lighting was appalling, reminding me of the lighting in this video. We were staying in a campsite on the outskirts of the city which we accessed via a 4 or 6 lane road with few or no pedestrian crossings. Driving that road in the dark was a worrying experience as pedestrians were crossing it from time to time. I came worryingly close to a woman with a toddler. The second night we passed the scene of an accident where someone had been knocked down and killed. Neither of us were surprised.

      The problem is that when you are in a badly lit area which contains bright point sources of light such as headlights of oncoming cars or the occasional street light, your eyes and cameras adjust to the bright lights and anything in the dark areas become invisible. Lidar, the other means of detection, works by bouncing light off the surroundings. If an object is mainly covered in something non reflective such as the unfortunate woman in dark clothes or is composed mainly of holes such as the bike, you don’t get a good return signal and so the object remains invisible.

      I suspect the unfortunate woman would have been hit if the car had been driven by a human.

    • #2582

      burp
      Participant

      I once missed an idiot who ran across an unlit highway mere metres in front of me somewhere north of Johannesburg. I was going at 130km/h and he was dark of skin and of clothing and emerged from the side of the road in a rural area.

      I have now watched that video and I’d like to claim that I’d have missed that pedestrian. 100% sure. Any human who couldn’t have avoided that accident shouldn’t be behind the wheel in my opinion.

      Sure, the pedestrian *was* being an idiot and wasn’t being particularly safety conscious. But idiots exist and lots of pedestrians act idiotically — the rules of the road say you’re not supposed to run them over.

    • #2583

      Aly
      Participant

      I have now watched that video and I’d like to claim that I’d have missed that pedestrian. 100% sure. Any human who couldn’t have avoided that accident shouldn’t be behind the wheel in my opinion.


      @burp
      Having watched that video I’m not so sure it would be that easy. I know the video quality isn’t great and one may have picked the pedestrian up sooner if they were actually in the car but they are crossing just into the shadow after the streetlights with glare from lights ahead as well. I make it about a second from when the feet first come into view until impact. Enough time to hopefully get on the brakes but not to stop without swerving or other evasive action. I’m surprised the car didn’t start to brake at least.

      I’ve just looked up stopping distances for various speeds (no I can’t remember them from my theory test!) and if dipped headlights give you about 50m of illumination, and something comes out of the dark at 60mph, even the most reactive driver will still be doing about 35mph when they hit it. The driving theory test questions say that dipped headlights illuminate only 30m ahead, you will barely have hit the brakes by then.

      Sure, the pedestrian *was* being an idiot and wasn’t being particularly safety conscious. But idiots exist and lots of pedestrians act idiotically — the rules of the road say you’re not supposed to run them over.

      Agreed

    • #2584

      toby
      Participant

      @burp I was nearly in a crash with a friend once, doing about 60mph, dual carriageway in daylight, twice the safe stopping distance to the car in front. There was a dog or something on the other carriageway, we both glanced across and back (split second) and the car in front stopped. I didn’t even see brake lights and he buried the brake pedal before realizing we weren’t going to stop, he took his feet of everything and steered. We missed by about 1 metre. The distance in that clip was a 1/4 to 1/5th, my feeling is that most people would have hit her.

      Cheers

      Toby

    • #2585

      Wod
      Participant

      I reckon the AI in the cars will take on the personality of the brand. So the tiny little white Google Waymo things will be safe to walk in front of because they will be driving really slowly so they can listen into your conversation, take pictures of who you are with and show you an advert as they go past.

      Ubers on the other hand will drive like mini-cab drivers, half their sensors will be there to look for cops and speed traps.

    • #2586

      girly
      Participant

      I’m assuming the vehicle was a completely silent EV?

      I suspect I do tend to step off kerbs initially if I can’t hear anything coming – I always look too, but I wonder if I step and then look – as the gutter shouldn’t have any vehicles in it, especially round these parts where traffic is uncommon and the roads wide. As a pedestrian I’m very conscious of sound to warn about vehicle behaviors (accelerating vs steady or braking etc) and possibly am more reliant on that than visibility initially. Clearly I don’t commit to crossing a road without looking but I can certainly see scenarios where I might get a foot off the pavement before looking. I really need to change that behavior….

    • #2587

      girly
      Participant

      Aren’t you supposed to drive at a speed that allows you to stop within the distance you can see to be clear e.g. the range of your headlamps? Therefore a driver driving correctly shouldn’t have hit the pedestrian.

    • #2588

      burp
      Participant

      @girly I agree! I watched the video and, from it, the pedestrian clearly did not just instantly appear where she was when she first showed up in the poor-quality recording. A human driver would have had more dynamic range and would either have seen her somewhere long before and slowed down, or have turned on brighter headlamps, or have slowed down because of judgement that the visibility was too poor to drive fast.

      I’m going with “I’d have seen her earlier and reacted quicker” but I guess the other two options are also possible.

      I think that any driver who did NOT do one of these three should not be behind the wheel.

    • #2589

      dopey
      Participant

      @girly You are right but almost nobody actually does, certainly not on roads like an unlit dual carriageway.

      How many times going up the A9 have you seen deer at the roadside, the first glimpse you get being as they flash past your headlight then your wing mirror even at modest speed? Particularly camouflaged things (often not intentionally camoflaged) really are very very hard to see in time by dipped beam headlights at any kind of speed. I know I’ve come horrifyingly close to a pedestrian wearing a brown wax jacket and tweed trousers walking a country lane on a dark night, I didn’t see her until she was basically alongside. I’ve also stopped within inches of a muddy black cow I couldn’t even identify until I’d backed up and moved the lights about a bit, I just had a feeling there was something there, both of those happened at around 30mph, they were low reflectivity, blended with the winter-colours background I expected to see and stationary with respect to it.

      If the Uber sensors are as it appears lidar and visible light cameras I suspect they could be susceptible to similar issues with dark low reflectivity materials. The unfortunate woman killed in the video may well have been moving at a speed and course relative to the car that produced little to no movement in the sensor’s field of view so if she was returning little to nothing for the lidar because of say for example a black velvet coat and a dusty/grimy bike she may just have looked like a slightly odd hole in the already distant background until the last fractions of a second. I guess in time they may publish their findings, they need the public to trust they’ve understood and resolved the issues.

    • #2590

      eric4
      Participant

      @burp I’m not sure the human eye would have a larger dynamic range than the cameras used. I say this because I’ve been peripherally involved in this and dynamic range was one of the challenging spec points set by the company we talked to. Dynamic range was greater than a human eye.

      Also note that in the UK I doubt you would have been able to cross a road like the one the accident occurred on. It would be regarded as too dangerous and barriers would be in place.

    • #2591

      Steve
      Participant

      @eric4 We can be pretty sure the human vision has a larger dynamic range than the ordinary screen and video format we see the video on.

      Difficult to put numbers for distance on it without going out in my car this evening to consciously judge visibility and distance but I would expect* to see a full sized adult already half way across the road (fully across the other carriageway and only moving at walking pace) before they entered “my side” of the road.

      *I’d be a fool to say I’ve never not seen something I should have seen.

      It doesn’t look like an urban road so in the UK you wouldn’t particularly expect barriers or zebra crossings at an ordinary junction.

    • #2592

      Newton
      Participant

      I don’t think there is any chance Lidar would not pick up an object that size and at that rage. My money’s on some weird AI failure.

    • #2593

      gordonbp
      Participant

      @newton I doubt there’s always a very clear line between what is and isn’t ‘detected’, there are so many competing system inputs it must be capable of handling significant uncertainty/noise from individual channels for operation in clutter. I just wonder if for some reason the lidar returns were degraded enough that less reliable inputs incapable of accurately detecting position because of the dark and relative movement because of the geometry were excessively prioritised. Presumably there’s a degree of risk of misidentified hazard it’s willing to accept to avoid too many false positive identifications which could lead to more accidents through out-braking or out-manoeuvring its human driven kin. It’s an odd one but you may be right, I suppose it could be one of those weird AI glitches where all the right stuff goes in in the right order yet something baffling comes out.

    • #2594

      Heyman
      Participant

      Tragic case, and shocking bit of video.

      Clearly something when massively wrong if the car didn’t ‘see’ that pedestrian. Also shows that humans will quickly loose interest in whats happening in front of them once most of the decisions are being made by the car.

      As far as I know, that is the second death involving self driving cars (the first being Tesla driver who slammed into the side of lorry).

      It would be interesting to know what total distance self driving cars have now covered, and how that stacks up with human drivers in terms of miles driven vs accidents/deaths?

      Purely a guess, but I’d be willing to bet self driving cars still come out safer, even at this early stage in their development.

Viewing 26 reply threads

You must be logged in to reply to this topic.

[snax_content]