How did LIDAR and IR not catch that? That seems like a pretty serious problem.
It's clear from the video that a human driver actually would've had more trouble since the pedestrian showed up in the field of view right before the collision, yet that's in the visible spectrum.
When I argue for automated driving (as a casual observer), I tell people about exactly this sort of stuff (a computer can look in 20 places at the same time, a human can't. a computer can see in the dark, a human can't).
Yet this crash proves that all the equipment in the world didn't catch a very obvious obstruction.
It's clear from the video that a human driver actually would've had more trouble since the pedestrian showed up in the field of view right before the collision, yet that's in the visible spectrum.
That's not at all clear to me. I don't know too much about cameras, but it looks to me like the camera is making the scene appear much darker than it actually is.
In the video, you can see many street lights projecting down onto the ground, and the person was walking the the gap between two streetlights. The gap between street lights (and hence the person) was in the field of view of the camera the entire time; they just weren't "visible" in the camera because of the low lighting. I'm confident my eyes are good enough that I would have been able to see this person at night in these lighting conditions. (Whether I could have reacted in time is another question.) It seems to me like the camera just doesn't have the dynamic range needed for driving in these low light conditions, which is a major problem.
I have to agree. Just like a normal camera has issues in low-light, it is clear that this camera is diminishing exactly how light the road ahead was. While I can't say confidently that I would have been able to stop to prevent hitting them, watching the video in full screen does lead me to believe that I would have seen them and been able to apply the brakes at least enough to reduce the impact. Also, watching the video of the interior it is clear the driver was looking at his phone or doing something else just prior to the impact. This alone leaves me skeptical to just how much could have been done to prevent this accident.
This is pretty much the experience I have with my dash cam, a Yi. In its recorded video, its automatic exposure control make it look like everything outside of the headline cone is pitch black, but it is actually not. I have seen deer and possums by the side of the road, and debris etc, that did not show up when I later checked the video for the same period. There is enough spillover light from modern headlights that a human whose eyes are dilated and adjusted to dark conditions will see a pedestrian standing on the median, stepping off it, crossing the inner lane towards the car's current lane. More than enough time to begin to brake and possibly swerve. I have dodged animals in a situation similar to this.
Yep that exposure control / sensor quality of the dash cam in the video was rubbish. My own Blackvues produce far, far better results than that. Just look at how nothing is illuminated by street lights, this clearly has the effect of making the poor rider appear "out of nowhere". Also agree it appeared driver was on smart phone most of the time, thus not in control of the vehicle, and had thus no business being on the road as these are systems UNDER TEST.
If that's the best Uber can produce then they ought to hang their heads in shame. Unless it was doctored with... as I find it hard to believe they'd put such rubbish quality cameras in their trials.
Do you trust Uber to provide all the data, or would they selectively produce data favorable to them?
Do you trust Uber to provide unedited raw video, or would they process it to increase contrast, make it appear that nothing was visible in the dark areas of the frame, reduce the resolution, drop frames, etc.?
The internal camera (let's be honest and call it the scapegoat camera, because that's the only practical use for human "safety drivers" when they are not permanently engaged) must take almost all its light from IR, because we don't see anything of the smartphone screen glare that the eye movement so clearly hints at.
I don't think the driver is looking at her smartphone. I think she's checking the car's monitor (as in a computer screen). Although to be fair, that should be showing the car's view of its surroundings so I don't know what's going on there.
Edit: Nevermind. Someone posted a picture of the car's interior, below and there's no computer screen.
Ok so this is getting old now, but I just came across the following - which show what I'd expect the roads to look like, and geesh were Uber ever full of crap to release their video which pretty much had the effect of exonerating them.
> I have seen deer and possums by the side of the road
Both of those have eyes that act as reflectors and you can see their eyes well before you can actually see the whole animal.
This[0] suggests that the total time required for a human to avoid an incident like this is 3.6s (at 35 mph, casual googling suggests the car was doing 40). Even if we add 1 second of extra time to deal with it I'm not sure that makes the cut.
Other people in the thread have pointed out the woman stepped out in a darker area between where the street lights are placed. Reflecting eyes are not the only way to detect an object. A person watching the road would have seen her dark silhouette contrasting to the next patch of light.
Also remember she was not a stationary object. She was in the act of crossing the road. Human eyes/brains are good at detecting motion in low light even if we can't 100% make out what the object is.
I have lived in Tempe and know that part of town well. There are apartments, gas stations, hotels, strip malls, fast food restaurants and a strip club. It's not a pitch black country road.
I know what you're talking about with the eyes, I spend a lot of time driving rural WA highways at night, but no. I have seen deer that had their heads facing the other way and were standing in the shoulder/ditch area. In conditions where i can definitely make out the shape of the deer and its location but the dash cam sensor misses it entirely.
Your last paragraph is a valid calculation if this were a case of a person stepping directly off a curb into the lane of traffic. However, it appears that they were probably standing on the median looking to cross, then stepped off into the left-most lane of traffic, an empty lane, proceeded across that lane towards the lane in which the car was traveling. In this sort of situation human intuition will recognize that a person standing on the median of a high-speed highway is likely to do something unusual. Particularly when you observe the visual profile of, as media has reported, a homeless person who is using the bicycle with numerous plastic bags hanging off it to collect recycling.
Driver didn't see this person because the driver was occupied with smartphone, only occasionally glancing up.
Also, has anyone here talked about the effect on the eyes of watching a (typically) bright white screen vs letting them adjust to the light of the night yet? This point deserves to be brought up.
Perhaps the video was intentionally darkened to simulate this effect. :P
>Also, has anyone here talked about the effect on the eyes of watching a (typically) bright white screen vs letting them adjust to the light of the night yet? This point deserves to be brought up.
Using bright interior lighting at night is something that we've known not to do for more than a century. If the driver couldn't be expected to see the pedestrian because the interior lighting or UX was too bright that is not something that does not reflect favorably upon Uber.
That's their only purpose. Nobody in their right mind could expect human observers to stay as alert as an actual driver when cruising for days with an AI that is good enough to not require interventions all the time. Passengers add nothing to safety, and an almost reliable AI will make anyone a passenger after a short while.
I'd like to have an interior view of what driver was actually looking at. It couldn't have been a FLIR monitor, for sure.. it seems more likely to be a phone held in the right hand? Bit hard to tell with the quality of the footage, but driver looked rather tired to boot.
If so (a hand held phone), in Australia that driver would be going to jail for culpable driving causing loss of life.
It could have been anything readable. I got the feeling it was either a Kindle or something like that or maybe even a hardcopy of something printed or written on paper. This was just a hunch but I think it's being validated in my mind by the fact that there was no light seeming to shine on the driver's face but that's probably due to the night vision camera not picking up that type of light? I don't really know. My mind is filling in a lot of gaps here, I realize.
EDIT: Upon re-watching the video a third time and really paying attention to this I don't think there is any real way for us to know without confirmation from the driver them self or an official report on the incident. My mind was definitely deciding things that just aren't discover-able from the video itself.
"Uber also developed an app, mounted on an iPad in the car’s middle console, for drivers to alert engineers to problems. Drivers could use the app anytime without shifting the car out of autonomous mode. Often, drivers would annotate data at a traffic light or a stop, but many did so while the car was moving"
The whole project seemed designed for an outcome like this. Eg allowing app to be used whilst on the move, after reducing from 2 to 1 operators. Culpability ought to lie with Uber.
I think he is, at least, I've never heard of any law that removes responsibility from a driver if driving a self-driving car. I think this will also apply to empty cars, if they get into an accident, the owner is liable.
I compare it to the backup camera in my car. While close up at night it is good, if something or someone is a short distance away I can barely make them out. However, looking in my mirrors I can see them or at least make out that someone or something is there.
A camera can have pretty good dynamic range at night, but it needs a big sensor, a huge lens to operate with a fast shutter speed. In the video, you can already see the motion blur, indicating shutter speed is slower than what it needs to be to identify nearby objects in low light.
Autonomous cars are never going to be viable. Just looking at the cost of high-end SLR sensors and lenses that you'd need to match human eye dynamic range, and you're already looking at an expensive setup, before we even get to things like 360-degree vision and IR/LIDAR/Hyperspectral imaging. And that's in addition to all the compute problems.
Sorry Silicon Valley tech-bros, but it's a fantasy you're chasing that's never going to happen. The quicker we can end this scam industry, the better.
I think you’re comparing object detection to high quality photography though. There are plenty of options that can detect objects at night. Even cheap infrared technology, I would think, would be sufficient for picking up moving objects at night.
Wetware is astonishing stuff. All the propaganda to anthropomorphize machines is showing here... cheap IR sensors are not the issue. AI is not intelligent and inanimate objects have no self.
They should pivit to augmenting drivers, not attempting to drive for them. I would happily utilize a properly designed HUD (meaning I have source access) connected to a fast MerCad or bolometer array.
Sorry for lack of input or varied discussion but I just had to stop and say how goddamn friggin cool it would be to have bolometers hooked up to a smart HUD that didn't interfere with your vision of the road. Something really translucent that smartly blended it's color scheme as to not interfere with the coloration of signs and details beyond your view on the road / around the road.
But you are right, though. I think augmenting drivers sounds like a great idea in the sense you talk about. The kind of augmenting drivers I don't want are those stupid headbands you'd wear that beep like crazy if your head starts tilting in a way that resembles falling asleep. If you are in danger of falling asleep at the wheel and need a device like that I think it's pretty obvious one should take a nap on the side of the road or in a free parking lot, haha. Hopefully if we do wind up headed in that direction the people inventing will have a similar way of thinking and inventing.
High-quality exists because the human eye is that sensitive and discerning. And there aren't plenty of options that can detect objects at night. IR isn't any cheaper, and then you have to figure out what IR bands you want to detect.
I've read (see [1]) that humans have a low-light ability that approximates ISO 60,000, a pretty large value and larger than simple video cameras provide. However, very high end pro/enthusiast SLR's go considerably higher, see this real-time astrophotography with the Sony a7s at ISO 409,600 (youtube video [2]). The same Sony will work great in full sunlight too.
The Canon ME20F-SH is a video camera reaches ISO 4,000,000. This camera has a dynamic range of 12 stops and is available at B&H for $20,000. [4]
Of course, this isn't exactly the challenge that cameras face when assessing a scene. The dynamic range happens within a single scene all at the same time. Wide dynamic range (WDR) is the term I've seen used in describing video cameras that can handle both bright and dim areas within the same scene.
No that's not how ISO works. The Canon ME20F-SH shoots high definition video at professional video shutter speeds and has an available ISO range of 800 to 4,560,000. At $20,000 I'm not suggesting that this exact camera would be appropriate for use in autonomous vehicles, but I am pointing out that video systems can now exceed the capabilities of human eyes.
There are a number of video samples shot on the Canon ME20F-SH on YouTube. In these one can see that under low light situations the camera is shooting at ordinary video speed (the camera supports shutter speeds from 24 to 60 fps). I'm not trying to push the Canon ME20F-SH; I don't have any association with Canon. The manual for this camera is available on-line if you'd like to read up on it: [1].
The actual exposure of a video frame or image depends upon the f-stop of the camera's lens (aperture), the shutter speed, and the ISO of the image sensor. See [2].
Basically, each doubling or halving the shutter speeds corresponds to one "full-stop" in photography. Each full stop of exposure doubles or halves the amount of light reaching the sensor. Changing the aperture of the camera's lens by full stops also doubles or halves the amount of light reaching the sensor. Full stops for camera lenses are designated as f1, f1.4, f2, f2.8, f4, f5.6, etc.
The light sensitivity of the film or sensor is also customarily measured in full stops. Very slow fine grained color film is ISO 50 and is usually used in full sunlight. ISO 100 is a bit more flexible and ISO 400 used to be considered a "fast" film for situations where more graininess would be acceptable in exchange for low light situations. Each doubling of ISO number corresponds to a full stop. So a photo take with ISO 400 at f2 with 1/1000 second shutter would have the same "brightness" as a picture taken at ISO 100 at f2.8 with 1/125 second shutter (less 2 stops ISO, less 1 stop aperture, and plus three stops shutter speed). Naturally, other factors come into play, the behavior of film or digital sensors at extremely slow or extremely fast shutter speeds isn't linear, there are color differences, and noise issues too. See [3] if you are interested in more about how photography works.
The footage from a normal camera should not matter, a self driving car is equipped with stuff that works regardless of light conditions like LIDAR oor IR cameras. This looks to me like a software failure.
The footage from the normal camera does matter in that it's the main way that we (humans) can process the scene. The parent comments are just pointing out that the camera footage is likely darker than the actual scene in person.
Waymo cars are capable of sensing vehicles and pedestrians at least half a block away in every directions. I was reserving any judgement on wether this collision could have been prevented, but seeing the video tells me that 1) a human driver might have hit the victim regardless, and 2) I'm very surprised that the LIDAR sensor didn't cause the car to stop to a halt much, much earlier. This is exactly the kind of situation that I would expect self-driving cars to be better than human drivers.
I agree that dashcam/external cam footage is going to be limited and possibly misleading, and I would think/hope such footage isn't the primary factor in evaluating accident cases. But I do think there's value to it. I shouldn't have said that it is the "main" way for us to process a scene, but the most accessible/relatable way.
What you posted looks pretty cool, I don't know enough about it to understand what I should be prioritizing focus on, but we can chalk that up to ignorance. The benefit that driver-view footage has is that it is a viewpoint all of us are familiar with. If you ask me to watch dashcam footage to assess some kind of traffic thing, there's a general expectation of where I keep my eyes and what I notice.
This normal-human-view mode is probably going to be necessary in AV cases in which we determine whether the car's AI did the right thing. Presumably, as AV becomes mainstream and extremely safe, these accidents will involve edge cases and outliers which are poorly interpreted by sensors/non-human-vision. Seeing the scene as a human driver does might be a necessary starting place?
But the Uber case in AZ, IMO, proves your point. The Tempe police quickly made a judgement call based on what seems to be inadequate video. Everyone who can now view the video will also be inclined to think how impossible it would be to avoid hitting the victim, even if the actual scene in-person has much more light. And of course, we don't want to judge AV solely on whether it performs as well as normal humans.
Uber may not be at fault, legally speaking. That's up to the legal authorities to decide.
However, as a society and civilization, and even more so, as engineers and scientists, we are going to expect that the autonomous car matches or exceeds human-level performance in critical situations like this.
Therefore the time spent on investigating, understanding, and discussing the root causes of the accident is worth understanding. Accidents like these generally do not happen due to a single factor. It is necessary to understand all the necessary factors if we want to make autonomous driving systems more reliable.
At the very least we need to understand whether the pedestrian appeared in the other sensors that a human could have identified by looking at the sensor data, and if yes, whether the autonomous system matched or exceeded human-level performance by detecting the pedestrian, and if the pedestrian was indeed detected, why the autonomous driving system failed to respond to the situation.
Surely not? Cars are routinely driven by people who are not owners, and liability for traffic offences (including that the vehicle must be insured) is with the driver.
In my experience typically only minor infractions like parking violations are assigned to the registered owner of the vehicle, but in other case – accidents, running red lights etc. – the driver is liable regardless of who owns the car.
No. The general rule is that negligence is required to be held responsible. If I let my next door neighbor borrow my car to go to the grocery store, and he hits someone, I'm not responsible. Unless, the person can prove "negligent entrustment", i.e. it was irresponsible just to let this person borrow my car, e.g. they're a habitual drunk, or blind, or 11.
However, most auto liability insurance covers whoever you permit to drive the vehicle, so the owners policy does typically cover the fender bender on the way to the grocery store.
Correct, the owner's insurance policy is the primary coverage when the owner lends their car to a 3rd party. Obviously in the case of a moving violation the driver is at fault and receives the penalty, but damage is still covered by the owner's policy. In the case where the other driver is at fault, that car's owner's insurance is liable.
exactly this. what's the response time of software? it ought to be close to zero and significantly faster than human's. let's say it's a generous 0.5s - no brakes where applied at all, and even with the crappy darkened video we got (place isn't that dark https://www.youtube.com/watch?v=1XOVxSCG8u0 ) the pedestrian was in view for 2 to 3 seconds.
car didn't see it at all even in those last moments.
Well it was a pedestrian but they were walking their bike across the road. It's not like the software should make a distinction between a cyclist in the way and a bicycle with no rider in the way.
Indeed, it's hard to find pedals without them. Even ones that cost $10 a pair have reflectors. Unfortunately, pedal reflectors are ineffective when the bicycle's path of travel is perpendicular to the light source. The video doesn't reveal evidence of other reflectors, such as the common spoke-mounted ones whose purpose it is to highlight a bicycle traveling crosswise. For a moment, the bicycle is clearly illuminated by the headlights; I don't see any spots of light on the wheels or elsewhere.
For a side view, the reflectors on the tires (visible at the end of the video) are way better indicators of “watch out! Bicycle” than those reflectors.
See this video for a comparison of visibility (not in English, but that's immaterial - set speed to 2x ;)): starting with a "bike ninja" and going all the way to "Christmas tree" https://youtu.be/oAFQ2pAnMFA?t=1m0s
It's from 2011, there's been a lot of improvement in consumer-grade cameras since. Even so, it fits my perception IRL: even a small reflector is orders of magnitude better than no reflector, and adding multiple (esp. covering 360 viewing angles) makes you stand out at night; same goes for pedestrians.
This 100%. When I drive, I watch the road. I don't watch my mobile phone, I don't watch the kids behind, I don't watch my wife. I don't watch the sky. I don't watch the GPS.
I just watch the road in front of me.
My idea is that the car has been behaving well for a long time and consequently the driver lowered is vigilance. Big mistake.
A fully attentive human driver might have hit this person regardless. Would they have hit them while taking no evasive action whatsoever? No swerving, no brakes?
I don't think so: the dash cam video is misleading. I had multiple ninjas jump at me before, and although I did notice and avoid them, they were not visible on the dashcam until the very last moment. Surely Uber would not release data to intentionally mislead the public?
Even so, I count a full second from when a human paying attention would have seen something just using this video as eyes, until impact. The stopping distance at 35mph is 136ft, which is 2.65 seconds at 35mph, so the accident would still happen but the impact speed could be lower.
Yeah, but at that speed, it's more than possible to swerve around an obstacle rather than screeching to a halt before touching it. Even turning slightly to the left/right would have made a dramatic difference in the outcome to this person's life. Not to mention the person in the car that might have also been severely injured if this was a heavier obstacle.
This was purely bad software, and no failure scenario being programmed in. I really don't think it's that difficult to program split-second reaction to obstacles that appear into the driving path. We need to get to a point where these vehicles can do stuff like this, even in a 2-dimensional way:
They seem to use 2.5 seconds as the standard for drivers to perceive and react to an obstacle, which based upon studies covers 90% of all drivers. 1.5 seconds to perceive, 1 second to react. Then you have maneuver time on top of that 2.5 seconds.
Given this, 1 second seems very low. A large percentage of drivers would probably plow into them at full speed.
This dashcam footage was released by the local police. It's likely they don't have the ability to access the autonomous car's working telemetry. Given Uber's legal history I doubt they'll release anything until they're compelled to by law. Personally I find it borderline irresponsible of Tempe PD to release this video and statements based on this video so early in the investigation.
Low beams at high speed do not give enough advance warning to reliably prevent a collision; as your lights are turned downward, you see a pedestrian only when they're quite close.
In general, traffic safety requires that road planners ensure that one of three conditions always applies:
a) the roads are lighted from above;
b) cars are able to use high beams;
c) there are no pedestrians crossing the highway.
This can be done in general, mostly by investments in infrastructure to ensure lighting or isolated highways wherever the density doesn't allow to drive with high beams.
> clear the driver was looking at his phone or doing something ...
Seriously, what else can you expect. These companies who do put these things on the road with the justification that "There is a human behind the wheel" should be taken out back and shot in the head...Just pull the plug. No more self driving cars for them. Those are just the kind of tech companies we don't want around...
See, it is not a mistake that they are making. They know well enough that this human behind this wheel is a useless as a dummy. But they do it any way. What does it say about them?
I feel sorry for the 'safety driver' here as it seems likely much of the liability will fall on her. As a transgendered ex-felon she can't have had a lot of fantastic job opportunities. I wonder how much Uber was paying her to sit in the hot seat.
The difference between Waymo and Uber here should be the difference between being allowed to continue, or getting barred from further self-driving research.
1. A driver who is not looking at the road cannot "potentially intervene", and is as good as no driver at all..
2. These companies seem to be doing nothing to make sure that the drivers will pay attention always and is always in a position to intervene. They even seemed to allow smart phone usage while they are in the car.
So, according to them, the human behind the wheel is just a decoy to prevent backlash from officials and the public, so that they can always say, "look, there is a human behind the wheel if something goes wrong"...
Also, even if they implement some measures, they can only make sure that the driver has eyes on the road. Not that they are actually paying attention. A driver who is actively driving the car will notice a lot more stuff than a passenger who is just looking at the road. There is no way to make a human pay that kind of attention with out actually driving the car. So at best, your "driver behind the wheel" is as good as a passive passenger.
And as told before, the companies are not even trying to make sure of that.
I could be wrong, but I believe part of the reason for having a human behind the wheel is that it allows the testing to take place under existing driving laws. At some point prior to an unmanned vehicle being allowed on the road, lawmakers need to have some kind of framework in place to deal with any incidents that arise. With a human behind the wheel, a fully autonomous car is legally no different to cruise control - it's just a driver assist, and the human behind the wheel is still ultimately responsible for whatever the vehicle does.
In that context, the landscape changes significantly - instead of a self driving car that mowed down a pedestrian, we have a driver who was too busy looking at her phone to pay attention to what her vehicle was doing. From the various articles, it seems that she's not an engineer, and is there in effectively the same capacity as any other Uber driver. If that's the case, she's putting far too much trust into an experimental system. I agree that Uber could do more in the way of technological means to ensure the driver is paying attention, but at some point, an adult with a job needs to be responsible for doing that job.
>lawmakers need to have some kind of framework in place to deal with any incidents that arise. With a human behind the wheel..
The framework should have been in place before these vehicles were ever put on the roads. For example, there should have been some formally specified tests for a self driving vehicle before it can be put be on the road, even with a back up driver..
> a fully autonomous car is legally no different to cruise control - it's just a driver assist, and the human behind the wheel is still ultimately responsible for whatever the vehicle does.
Any thing that does not require drivers to keep their hands on the wheel is not a driver assist. It IS the driver. So there should be tests that make sure of the competence of the tech that is in the drivers seat.
>they can only make sure that the driver has eyes on the road. Not that they are actually paying attention. //
I'm certain that if you can design and build a self-driving car that you can design a simplistic human attention monitoring system that will cause the car to pull over if attention level is too low.
Gaze monitoring that checks for looking downwards or away from the carriageway for extended or too often repeated periods wouldd probably be enough.
I imagine the attention of the "vehicle operator" is vital to the proper training of the vehicles -- if they don't see near misses, or failures to slow for potential hazards, or failures to react to other road users then how can the softwares faults be corrected? Do they get a human to review all footage after the drive?
I agree completely. As far as I can tell, the driver did not even have hands on the steering wheel. How hard would it have been to put sensors on the steering wheel to require both hands? They didn't even do that. Although even if they did, I agree with your statement that "[t]here is no way to make a human pay that kind of attention with out actually driving the car."
Not difficult at all, and you can make them keep reasonable attention. Look at the new Cadillac driver assist: sensors in the wheel for hand placement -and- eye tracking. If the driver isn’t watching the road/holding the wheel, they get escalating alarms until the autopilot disengages.
And that’s consumer drive assist tech, not “we are experimenting with full autopilot” tech, where I’d think such safety measures would be even more appropriate.
This is a solvable and solved technical challenge. Uber just didn’t devote any resources to it because they don’t appear to give a shit beyond acquiring a legal fig leaf to shift liability from themselves to an individual.
Frequent, randomly scheduled disengagements should keep the driver quite on edge, preventing them from becoming a passenger. But each and every one of them would create additional risk, so the net improvement might be negative. There is just no way to get this right, except for being reluctant of pushing to scale. With all the hype, wishful thinking and investor pressure, this clearly isn't happening.
I've been thinking about this for the last couple of days, and it's definitely a hard problem -- even with steering wheel sensors and eye tracking, it doesn't stop people zoning out and not being ready to react.
I did wonder if you could require the driver to make control inputs that aren't actually used to control the car but are monitored for being reasonably close to how the computer is controlling the car, and then the automation disengages (with a warning) if the driver is not paying sufficient attention. I then realised that may be _worse_ - in the event of a problem, the driver would have to switch to real inputs that override, which may delay action and not be something they do automatically. It would mean they are paying attention more to see if the automation is making errors where they have more time to react though (e.g. sensor failure that is causing erratic behaviour but not led to an emergency situation).
I wonder if a hybrid approach might be viable -- fake steering is used to ensure that the driver is alert and an active participant, but the driver hitting the brakes immediately takes effect and disengages the automation.
> Also, watching the video of the interior it is clear the driver was looking at his phone or doing something else just prior to the impact. This alone leaves me skeptical to just how much could have been done to prevent this accident.
Wait, aren't you mean to have your hands on the wheels at all times? I don't see what to be skeptical about when if he just followed the law this could have been avoided.
It seems to me the driver might be in for some legal trouble.
But this has got to be just the black-box camera, right? Surely the actual camera they use as a driving sensor is much better than this? Not to mention the LIDAR and all the other sensors that should have caught this.
> Also, watching the video of the interior it is clear the driver was looking at his phone or doing something else.
Probably checking the computer installed for diagnostics of the autopilot system. If it's in self driving mode and you are the engineer in charge, you'd want to constantly check what the system is seeing vs the actual conditions on the road.
If you're the driver of a car you're supposed to ensure safety by looking out, not verifying sensory information. If Uber designed their cars to show a rendering of the computer's perception to the driver, or other sensory output, they would violate that principle.
To me it looks like the guy is just falling asleep at a boring job. In all likelihood that was not an engineer more than any other taxi driver is an engineer.
The software is the "driver" of this car. Not the human behind the wheel. Take a look at job descriptions [0] for this. They always include a bit about "operating in vehicle computers". The fact, we don't know what the person is doing.
I am pretty sure Uber uses an iPad app for its autonomous vehicles. The driver is looking at that iPad application periodically along with the physical windshield view.
If you search "Uber autonomous vehicle" you can see some videos of the display. From what I gather, basically gathers the signals into a human readable model. In general I wouldn't have recommended this driving style but it might have been too dark to see much anyway.
I don’t understand this, I’ve seen a few people comment in the same vein.
People can safely drive in total darkness with the aid of their amazing human eyes and high-beams.
If for some other reason visibility is low you slow down - not rely on glancing at a backlit display ruining your own night vision and taking your eyes off the road for seconds at a time.
Or flick your high beams, quick beeps, adjust speed... I do all these things if I see anything on a collision course with my vehicle.
It is surprising to learn that these vehicles are operating at night. To collect training data, since nighttime driving is inevitable, perhaps there are ways to simulate night to the computer vision systems during daytime so the human supervisor can still see clearly.
Lol, the "human supervisor", looking at his knees, probably on reddit or tweeting.
Would you trust this system that didn't even manage to slow down at all with a pedestrian slowly pushing a bike directly in front of it, artificially adjusted to be even worse, driving during the day??
Human eyes have the same issue: if you are next to a bright light source, the areas without or less light will look much more dark. I assume cameras work the same way?
Cameras work the same way, but much much more poorly. A human eye can see multiple orders of magnitude higher range of light to dark areas at the same time. The accepted estimate is that the human eye can detect a 1 million: 1 range from light to dark in terms of photon intensity.
The driver has been described as male in news reports:
> "The driver said it was like a flash, the person walked out in front of them," Moir said, referring to the backup driver who was behind the wheel but not operating the vehicle. "His first alert to the collision was the sound of the collision."
also note that pushing the brakes was not the only option : steering the wheel to avoid collision was another, maybe more efficient. Still, I feel the same as you do : I cannot guarantee I would have avoided this.
I can confidently assert that Asian or atleast Indian drivers will almost assuredly not hit the pedestrian in this scenario; We have trained our eyes and senses to watch out for these as it happens all the time.
EDIT: what i meant, in light of the downvotes is that humans can train themselves to see, and just that folks driving in Asia have heightened sense of alertness, due to their environment. Hope it came out alright.
The whole reason many people on here have been advocating for self driving cars is that they can see obstructions more or less perfectly in the dark with LIDAR. I am much more interested in what that sensor said.
I'm reluctant to infer exactly what a human eye would have seen in that situation. I have absolutely driven down streets in suburbia where the gap between street lights was large enough to make them quite dark, and that video was an example of exactly what I was afraid of happening whenever I drove down those streets (though admittedly my fear was hitting a white tailed deer).
I think it might also be fair to argue that the car's high beams were not on (but again, that shouldn't matter because of LIDAR, right?).
I'm not confident even an above average human driver would be able to avoid that accident, even if good eyesight gave you an extra half second to respond. Dark clothing and no reflectors means that person was definitely invisible to both the camera and the driver for some time after they would have been visible in daylight.
I've had a couple of situations where someone appeared close to my line of travel with low visibility clothing (at night) that scared the living shit out of me, and they weren't trying to cross the street.
To be clear, I am not blaming the victim here, but do wear high visibility clothing when you're a pedestrian near high speed roads at night.
A person with common sense and a developed understanding of the situation would drive more slowly in situations like this. The law says that you don't drive faster than you can see.
A similar thing (no fatalities, just a shopping cart pushed by homeless people) happened to me. Ever since then, I have learned to be much more aware of situations like this (tunnel of light surrounded by darkness).
This just shows that Uber's tech is bad and that they let it on the road shows that their culture is still at least partly rotten.
I don't think Dara is the do gooder that some people are making him out to be. His primary motivation seems to be to usher Uber to an IPO. IMHO, if he actually had ethics, he would be front and center on this. Your company just killed someone. Where are you?
> The law says that you don't drive faster than you can see.
Amusingly, the law also says that manufacturers have to produce headlights that cast light out far enough to leave you adequate stopping distance at 60mph. Almost no headlights on the market currently do that.
Not a counterpoint, just a tangent that I find sadly amusing.
The ninja is the reference standard for real-world pedestrians. It's up there with the surprise moose. Systems that can only detect bright peds are going to be horrific meat grinders and lead to autocar hell instead of autocar heaven.
Sad side note that most people appear unaware of the benefits even the simplest and cheapest of reflectors do provide.
The seemingly random design decision of many runner manufacturers to embed tiny reflector strips in their shoes have no doubt saved countless lives. And their owners would probably be none the wiser.
And those who are aware often lack the understanding that with glare-minimizing headlights, reflective surfaces at or below knee-level are many times more useful than anything higher. A reflective hat would be pointless.
Yup. There's a place for education here. I know I wasn't really aware of the benefits until adult age, when I started to find myself more often in a car, at night, in rural areas. I still remember the first experience, in which I've noticed a cyclist on another lane ~0.5 seconds before we passed him. Dark clothes, dark bike, zero reflective elements.
In Norway, when growing up, I was frequently exposed to campaigns saying "Bruk refleks!" (Use reflector(s)!), and given free ones at every opportunity.
Of course it makes sense there where daylight may be hard to find half the year, however even in Australia, once it is dark the darkness is the same.
And I haven't seen a single government initiative to increase visibility awareness - most people are completely in the dark. (Sorry)
Riding shared bike trails in Melbourne at night on the commute home, this is something I think about often in the "winter" months. Peds may hate the strong glare from my LEDs, but it is the only thing the has half the chance of making out ninjas against the frequent sports ground stadium floodlights the path goes by.
> The whole reason many people on here have been advocating for self driving cars is that they can see obstructions more or less perfectly in the dark with LIDAR. I am much more interested in what that sensor said.
That is not the whole reason, it is one of many reasons.
> To be clear, I am not blaming the victim here, but do wear high visibility clothing when you're a pedestrian near high speed roads at night.
Preventing the accident might not have been possible, but even being able to decrease speed by a tiny amount would have greatly improved the pedestrian's chance of survival. Slowing from the 38mph that the car was traveling down to 30mph would decrease the chance of fatality from about 45% to below 10%.
Yep. I live half a mile away and just drove the same path tonight around 10pm, it's nothing like the video. There are spots that are darker than others, but they don't look nearly as dark. Nowhere on the street looks pitch black, there's ambient light everywhere.
For anyone who's interested, try taking your phone with the camera app open into a dark room and comparing what you see to what's on the screen. Which shows more detail?
Out of interest-- can you take a pic while the lighting outside is similar (assuming weather hasn't changed dramatically?) and maybe adjust exposure to what your eyes see? Or take a camera phone pic for comparison?
> The gap between street lights (and hence the person) was in the field of view of the camera the entire time
But the gap between street lights is going to be very hard to see into.
> I'm confident my eyes are good enough that I would have been able to see this person at night in these lighting conditions.
I think you're overconfident. Human low light vision is very good if there is low light everywhere. But it is not good at seeing into low light regions when brightly lit regions are nearby.
That said, I agree that a visible light video camera is likely to be even worse that human vision under the given circumstances. But as others have commented elsewhere in this thread, the car is not just supposed to be using a visible light video camera. It has LIDAR and IR sensors, which should have clearly shown the pedestrian well before visible light did.
FWIW the spot where the crash happened is in fact badly lit. I know this anecdotally from having been at the location for events -- it's right next to a concert venue -- but it can also be seen on other dashcam videos.
In this video [1] driving northbound, same as the vehicle in the crash, the car first goes under AZ-202, emerges under a streetlight, goes through a darker spot, then another streetlight (as you see the rocky outcrop), and then a very dark spot: and suddenly, you see a right-turn lane that wasn't there before. The latter dark spot is where the crash happened.
Another video by the same author, driving southbound [2], provides another useful reference. And these videos are three years old, yet the illumination of the roadway has not improved. Cameras exaggerate the contrast a bit, but not unreasonably so. The streetlights in question essentially aim directly downwards, illuminating the roadway immediately underneath, but much less of the surrounding air than other designs. This is responsible for the dark gaps, albeit it does significantly reduce light pollution.
Found more. The car in this video is going southbound, camera facing backwards [3]. This view faces the same way as the Uber did, but of course this video is moving away from the scene, and offset by a few dozen meters to the west. The drastic change in roadway illumination can still be seen.
In a fourth video [4], the car is going northbound, like the Uber, in the proper lanes, but the camera is pointing obliquely front-right. The illumination seems better, but you can still see the intensity of the shadows, including environmental shadows and the car's own shadow, as it moves between the lights.
Everyone is moaning and slicing and dicing what the self-driving vehicle did wrong but, since you're familiar with the area: are pedestrians typically expected to be crossing this road?
Seems like the accident has a lot of factors that might not only be the self-driving car's fault, nor even a human driver that was fully in control. Regardless of how well people may want self-driving cars to do, one thing that can actually exist in the present is to make sure that we are creating safe ways for pedestrians to cross a road.
I've also driven around here a lot. No, pedestrians are not common. Maybe once a week in my experience? They do love to cross outside of crosswalks at night, though, and I've found that I have to adjust my own eyes' object recognition to look for moving shadows and not just moving lights, because they're very hard to see even in well-lit areas.
I've driven many thousands of hours at night and have dealt with a fair number of crazy pedestrians including a rather ... uncoordinated ... guy in Casa Grande who decided to go in circles on his bike in the middle of the road at around 3 AM for no discernible reason. Fortunately that place was much better lit and I was able to see him and stop until he got out of my side of the road.
So it's not that common, but yes, every so often you will see some person in black jaywalking across a wide road at night and they're quite hard to see. I don't think a lot of people appreciate that the streets here are wide & fast and that there just isn't that much pedestrian traffic even in daytime.
That was my suspicion. I've lived in very suburban areas before as well as rural ones where you might even be going 55 on a two-lane road with no street lighting whatsoever.
Here in LA, it's dense and traffic can't get up to very high speeds and we have relatively frequent places to cross safely if people choose to do so. I've definitely seen those who choose not to walk an extra 100 feet to wait at a crosswalk nearly hit in dusk or night traffic.
No amount of automation is going to bring the accident rate down to 0 so through a combination of factors, such as traffic and community design, we can work in tandem with automated driving to get closer. There's still the X factor of our human ability to do really dumb stuff.
Tucson does it due to the nearby observatory. The greater Phoenix area has a huge glow that washes out all the stars. You can see the glow as far away as Casa Grande when you come out of the little rocky pass on I-10 north of there.
> The greater Phoenix area has a huge glow that washes out all the stars.
I live in the Phoenix area, on the west side closer to Glendale (specifically, the border between Phoenix and Glendale is literally in my back yard).
There are times in the summer where the glow from the city is so bright, that rather than a dark sky (never black), you have a grey dimly lit sky instead.
Literally, "the sky was the color of television tuned to a dead channel" - maybe not as bright as the static Gibson was referring to, but still bright enough to see by - even without a full moon.
This site lies on the approach route for Sky Harbor airport. I'd imagine the street lights are intentionally designed to reduce light pollution at the expense of "on the ground" effects.
> But the gap between street lights is going to be very hard to see into.
This wholly contradicts my experience driving at night on a street with street lights. I can't recall a time in my entire life I have had significant difficulty seeing into the gap between street lights. Keep in mind that the gap is not arbitrarily chosen.
Edited to note that I have experienced difficulties in low-vis conditions such as snow storms, sand storms, VERY strong rain storms, etc. None of which apply to this situation.
Keep in mind a lot of folks in this thread might be suddenly realizing they have reduced ocular ability at night, a likely common condition that pretty much nobody is aware of when it’s minor (because it’s not obvious something is amiss; maybe it’s just that dark). I agree with you that streetlights and headlights are almost universally sufficient in my experience. If they’re not, it’s worth getting your eyes checked out for light sensitivity at night. You never know.
I’m not sure a typical eye exam checks for it, either, because none of the tests I can think of seem like they’d be useful.
(As usual, an even keeled comment based on family experience is -2 and rapidly being silenced with zero feedback inside 5 minutes, which makes me wonder why I contribute to this community at all, probably time to stop)
I did one at my last eye exam and it was pressing a button when you see dim flashes in all different locations. If you had low sensitivity, you wouldn't see those flashes and presumably you'd get a low score.
That test mostly isn't testing sensitivity, it's testing field of view which is an indicator of some potential eye health issues. It might end up testing sensitivity incidentally but that's not the purpose.
I suspect it's too late to chnage now but have a "throwaway" account is an indicator you might not be committed to the community. one has to dig a little deeper to find out a multi year history with 4000 karma. so first impressions of your comments might be getting biased (it might just the "red car effect" but i am seeing a lot more throwaway accounts these days)
I would also not judge the community based on reactions to this very contentious thread - i am wary of jumping in on this one, but thought it worth noting your comment was not wildly out of place.
Judging a comment, or commenter, based on karma is asinine. Respond to comment, not commenter; ideas not people. You are not well representing HN, and this is my 'unpopular opinion' account talking. There are plenty of better ways to engage, and I do appreciate your enthusiasm for HN. Perhaps this is an apt introduction to the heated discussion that is HN.
Well, in my view I was responding both to the comment ("i am leaving") and the commentor (making the years of participation and 4000 points relevant). If someone has been a contributor for many years then we should consider why they chose to leave. it might be them, it might be us.
I actually believe i am representing HN as a place where different opinions can be voiced, hopefully in a manner to generate light not heat. Heated discussions are rarely the useful or interesting ones to read.
Thank you for appreciating my enthusiasm.
PS
Are you using two accounts - one ("my 'unpopular opinion' account) for saying things you fear people might not like? That seems odd. May I ask why?
This may be how you believe you see the world but most people take reputation into consideration and on sites which expose that information account age and karma are very popular cues for that.
Karma does not mean shit. It just means you are complaisant. I think the proper way to use things like HN/reddit is to always use a throwaway account and always speak your mind without the fear of negative karma...So I also agree 100% with the parent. Reply the comment, not the commenters, their karma or their entire history.
> Keep in mind that the gap is not arbitrarily chosen.
It's not supposed to be, no. But the gaps are not always optimal. The spacing of the street lights in the video (to the extent I can tell) seems to be quite wide, wider than I would think is optimal.
The edges of the "lightpool" that the lamps normally cast is probly being clipped by the cameras crappy dynamic contrast, it is almost certainly a much larger lightpool in real life.
If Tempe is like Tuscon, they are using different kinds of street lighting from the rest of the country to minimize light pollution for star gazing reasons.
> But the gap between street lights is going to be very hard to see into.
Looking, right now, at a parking lot between two lights from a well-lit room. I can make out most of the outline of the black car in the middle of the "darkness" without any trouble. This isn't even the low light vision kicking in (which I agree isn't going to kick in if you're driving). Human vision should be able to make out the pedestrian earlier than the video footage.
How long does it take you to make out the black car and determine that it's a car? What if the car were coming straight at you out of darkness and you were standing in the light of a street lamp?
Also, are you looking straight at the car? Or are you looking elsewhere so that the car is in your peripheral vision, the way it would be if it were on the side of a road you were driving on?
Not when you transition from high to low light conditions. The problem is night vision has more noise which makes movement detection far more difficult. This is made worse because the pupil can't fully dilate making the gaps seem much darker.
This street in particular is weird at night because the street downstream rises up, and the light from those lamps is cast at a higher point. The place she was hit is extremely dangerous because there are no lights on her, and no lights behind her.
I believe that a human would be able to see in those conditions. It's a lit street with a car with functioning headlamps. It wasn't foggy or rainy.
I've personally driven down country roads without any lighting except my headlights and saw deer poking their head out of the woods a ways away for which I slowed down in case they darted across the street. Someone slowly walking their bike would be trivial.
The video makes it seem impossible but afterwards in the interviews the driver said it wasn't too bad after his eyes adjusted. He did have some issues with his own lamp blinding him which lead to errors. (He actually won this stage.)
As far as I'm concerned Uber's software/hardware is completely at fault and not ready for public testing. I'm uncertain how much better everyone else's tech is but Uber's typical carefree approach has ruined it for everyone.
There are consumer level dashcam that can shift up to 12800 ISO which can create a fairly distringuishable picture with ambient moonlight.[1]
Canon builds sensors with ISO's in the millions which should be able to see distinguishable shapes without ANY light. [2]
> It's a lit street with a car with functioning headlamps.
The headlamps may have been functioning, but they appeared to be aimed way too low. You can see that the car is able to traverse the distance lit up by the headlamp in about a second at 38 mph. If the headlamps were aimed properly, it should light up the road about 5 seconds ahead of the car.
> If the headlamps were aimed properly, it should light up the road about 5 seconds ahead of the car.
A system with this rule baked in would be driving slower.
People adjust the way they drive based on what their environment is doing, how well their equipment is working and their own alertness. Except in the extremes we should not accept misconfigured equipment as an excuse. And if a system detects that there is no acceptably safe speed for it to go then it should not move at all.
> A system with this rule baked in would be driving slower.
Arguably, the system should detect a misconfiguration like this when the car is turned on and not allow the car to be driven until the problem is fixed.
I also suspect that human eyeballs would have a different view of the light/dark portions of what's depicted there, and especially eyeballs would have probably had a much higher chance of detecting movement in peripheral vision than that video gives any hint of.
We typically cant see much detail in the scene out of our small region of focus, but you can bet if a tiger appears from behind a tree our visual system will scream to the brain _look over there right now!_
Our eyes and our entire visual processing system is very much not "just like a webcam, but made out of meat".
I’ve driven that location on that road many many times at night and no it is not that dark, it is lit up well like most city streets. The video make the contrast appear greater.
I have a dashcam, and I've seen night videos from it.
In fact, the picture from my dashcam seems much better than this low quality mess, but still night videos from it come out much less visible than reality.
I've tried to rewatch some parts of videos later, and I find I was able to see much more detail on the sidewalk and on the periphery than was captured by the dashcam. Everything gets blown out in the night videos by the headlights.
Which makes me wonder if the Uber autocar is just relying on camera vision to drive itself... If it is, and it's been lying to authorities (I don't know what they said to the authorities about their cars' capabilities), that could be big.
The point is not germane. What is germane is that a car that supposedly uses LIDAR and infrared, and presumably was approved by the regulators on the basis of such, should have had no problem seeing the pedestrian as LIDAR and infrared are unaffected by night and at least shown some indication of braking but did not. This suggests that the car does not in fact utilize any of those fancy (non-vis) detection methods. Alternatively, these fancy detection methods were fooled by the bicycle and thus misclassified as an error or something.
My point is that it's likely the camera view we're seeing has nothing to do with the self driving portion of the vehicle (a good hint to this is the interior view--useless to autonomous driving, but a common feature of dash cams).
The car has a LIDAR sensor mounted on the roof. It is supposed to continuously scan 360° of the environment. Since LIDAR is an active sensor (it emits light), the car should have seen the person and bicycle even in the dark. That it did not do so suggests the car does not evaluate LIDAR input, or it dismissed the object as erroneous data.
It has 64 lasers, spread out over about 27 degrees - about 0.4 degrees per laser, from almost horizontal to an angle of 24 degrees or so down. Now take a look at where it is mounted on the car, and envision these laser beams spreading out and being spun in a circular conical area around the car.
Now - if you think about it - as the distance from the sensor increases, the beams are spread further apart. I'd be willing to bet that at about 200 feet or so away from the car, very few of the beams would hit a person and reflect back. Also - take a look at the reflectance data in the spec. Not bad...but imagine you are wearing a fuzzy black jacket on your top half. How much reflectance now?
What do you think the point cloud returned to the car is going to look like? Will it look like a human? Hard to say - but you feed that into a classifier algorithm, there's a possibility that it's not going to identify the blob as a "human" to slow down. Especially when you add some bags, a strange gait, plus the bicycle behind the person. All of this uncertainty adds up.
I am also willing to bet that only the LIDAR was used for collision detection (beyond the radar on the unit). Any cameras - even IR based - would likely only be used for lane keeping and following purposes, plus traffic sign identification. Maybe even "rear view of vehicle" detection. Ideally it would be used for "person/animal" identification and classification to - but again, given the camera sensor, and who knows what the IR sensor saw or didn't see, along with the weird lighting conditions - well, who knows how it would have classified that mix?
Lots of variables here - lots of "ifs" too. All we can do is speculate, because we don't have the raw data. Uber would do well to release the entire raw dataset from all the sensors to the community and others to look over and learn from.
Finally - I am not an expert on any of this; my only "qualifications" on this subject is having taken and passed a couple of Udacity MOOCs - specifically the "Self-Driving Car Engineer Nanodegree" program (2016-2017), and their "CS373" course (2012). Both courses were very enlightening and educational, but could only really be considered an introduction to this kind of tech.
> The dynamic range of the human eye is vastly better than a visible spectrum camera.
Certainly better than any camera mounted on a dashboard.
It's honestly a bit surreal how the pedestrian appears out of the splotch of pure darkness in the frame. That's low dynamic range and resolution (or high compression) at work, not how light behaves in reality.
I figured that light in front of the car was mostly just messing with the camera but that driver sure didn’t see that pedestrian either. I’m willing to give a human driver the benefit of the doubt here and say that even with eyes on the road and hands on the wheel the outcome would likely have been the same. The pedestrian was not highly visible - no reflectors, dark clothes, it’s really hard to see people like this.
The eye can gain a lot more stops through adaptation (irising, low-light rod-only vision), but those mechanisms dont come into play when viewing a single scene -- and cameras can also make adjustments, e.g. shutter speed and aperture - to gain as much, if not more, range.
A camera captures the entire scene in a frame with a fixed dynamic range. Human vision builds the scene with spatially variant mapping, the scene is made from many frames with different exposures stacked together in real time.
I'm concerned about poor scotopic adaptation due to the rather bright light source inside the car - maybe it's the display he's looking at. I see a prominent amount of light on the ceiling all the way to the back of the car and right on his face. It's really straight forward to collect the actual scene luminances from this particular car interior and exterior in this location, but my estimation is the interior luminance is a bigger problem for adaptation than the street lights because the display he's presumably looking at has a much wider field of view, and he's looking directly at it for a prolonged period of time. It's possible he's not even scotopically adapted because of this.
And also why is he even looking at the screen? He's obviously distracted by something. Is this required for testing? Ostensibly he's supposed to drive the car first. Is this display standard equipment? Or is it unique to it being an Uber? Or is it an entertainment device?
Retest with an OEM lit interior whose driver is paying attention. We already know the autonomous setup failed. But barriers are in place than also increase the potential for the human backup driver to fail.
I agree, but I don’t think the eye can adapt beyond its inherent dynamic range over a matter of milliseconds - the iris is not opening or closing over that timescale, so you’re relying on the inherent dynamic range of the retina (which is pretty good).
What the eye IS doing is some kind of HDR processing, which is much better than the gamma and levels applied to that video. I bet a professional colorist could grade that footage to make it a much better reflection of what the driver could see in the shadows - even with a crappy camera, you can usually pull out quite a bit of shadow detail.
It's not so simple. Technically, you are not wrong but a video feed should have been sufficient here. It should also be considered that digital video has improved drastically over the last decade.
Even LIDAR aside, computer vision and a raw video feed should have been enough to have prevented this collision.
When a digital camera records an image, a gamma curve is applied to it before display, which makes up for our bias against the darker portions which the digital equipment does not have. We are very capable of guessing the results of bright conditions but not dark conditions via compressed video.
Moreso, these cars should not be using consumer CCDs with compression. They should be utilizing the full possible scope of video.
> When a digital camera records an image, a gamma curve is applied to it before display, which makes up for our bias against the darker portions which the digital equipment does not have.
Gamma correction makes up for a bias against darker portions in the display, not in our eyes. It's a holdover from the CRT days where the change in brightness between pixel values of, say, 10 and 11, was far less than the change between 250 and 251. Human eyes have excellent low-light discernment which is why 'black' doesn't really look black and you can make out blocky shapes during dark scenes on some DVDs.
Compressed video lacks information in the blacks and that is why we see blocks. The blocks are not there before compression, so it’s not simply a matter of detecting them. While we are good at seeing objects in blacks, your explanation alone doesn’t account for why compression algorithms reason to remove so much of that data. Maybe we are saying the same thing. It’s hard to tell.
Your assertion about the origins, however, are at odds with what I have been taught, my understanding, and all the supporting info I am finding in a quick search. My understanding is that luminance values from a sensor have something of an empirical scale but I’m sure this no complete explanation. I am speaking from my working knowledge. I can’t find anything supporting that it is simply a fix for discrepancies between display types. Can you link to something or explain what I am missing?
I do know a bit about cameras, and you're spot on.
You will frequently see dash cam footage and night photography blow out the relative highlights and blacken the relative shadows.
This is because (cheap) hardware does not have the same dynamic range as human eyes, especially at night. So "properly exposed" it has to make a call to capture light values in the middle somewhere. Those light values too far out the top it interpreted as white, those out the bottom it interpreted as black, created an artificial high contrast version of what a human eye would see.
This is pretty intuitive, generally when we're driving down the road with our lights on, we aren't literally moving between pools of black, often in many urban areas I'll even forget to turn my lights on because I can see well enough.
You MAY be able to get a VERY BAD interpretation post processing of what a human would see by increasing the brightness of those pixels near the black threshold.
If any of you have a dash cam it's very obvious how the light levels of images captured at night look like this video and is VERY different from what you see as a driver - objects are much brighter than this with your own eyes.
Also - the car is driving way too fast.
I did some driving tonight and paid close attention to when I naturally slowed down - and albeit I'm probably on the higher curve of good drivers in that I don't tailgate, drive the speed limit and generally slow much slower than the speed limit when conditions are poor (fog/rain/snow, night, slick/wet roads, near curves/hills where I can't see the road). I noticed that many of the times I naturally slowed down on the roads here I slowed considerably under the speed limit by 10 to 20 MPH in some areas. It seems this Uber SDV is generally going as fast as it is possibly allowed to regardless of what it can see.
Also even if the camera was perfectly accurate about human field of view, no human driver in his right mind would drive so fast with such a poor visibility. Any judge would qualify this as reckless driving.
So either way the software failed:
-If AI misjudged Lidar information and didn’t compute the slow moving pedestrian it’s a fail
-If it didn’t have enough computer vision space it should have slowed down
Possibly in the second scenario the human test driver is at fault too because he should have noticed bad condition and hit the autopilot kill switch.
In France it’s 100% (in civil cases), unless the driver can prove it’s a suicide. It just goes by kinetic energy: you store it, you are responsible for it. Other people don’t have to dodge your car. And since death penalty is not part of the arsenal, killing a pedestrian is not an appropriate sentence if they commit a infraction that’s punished with a 50€ fine.
I don't really know how it works in the US or in this state, but in my country, you simply can't drive when it's as dark as the video appears. Either you're not in a city and you can turn your mainbeam headlights on (the blinding ones), or you're in a city and the road is much more lit and the speed limit is 50kph.
With those, the driver would've seen her from a mile away.
Since in the video, we aren't seeing the original scene, but rather, the camera's interpretation of that scene, I think it would be hard to judge except to base it on what your average streetlight brightness is.
Like a camera, your eye also has only so much dynamic range. So if those street lights are bright enough, or your interior lights are too bright, you might have nearly zero visibility in those shadows.
But it is certain that a self driving car "should" be able to see. Even two cheap digital cameras one tuned to see the darker range and the other brighter should easily see in these type situations.
"Even two cheap digital cameras one tuned to see the darker range and the other brighter"
Sounds a lot like rods and cones in our eyes, huh?
There's another difference with eyeballs that would almost certainly have helped here - the low light sensitive peripheral vision that the rods provide is also attuned to movement, we're more sensitive to movement in peripheral vision as well as being better able to see in low light.
You wouldn't need two different cameras, just one camera shooting HDR video (alternating between over/underexposing the frame so that no information is clipped) to get a clear image at all exposure levels.
Eyeballs are pretty good at night vision once adjusted, but good high sensitivity cameras can be much better. And let's not get started on LIDAR/RADAR... it seems clear to me that this was not a sensory deficiency, it was poorly designed/tested software.
No, much more like the iris in our eye. Turn on a bright light inside a car and see how hard it is to see outside on a dark night. Modern HDR cameras have a much higher dynamic range than the human eye. Hence the surreal HDR photos you see.
The driver wasn't fixed on the road but he glanced two times prior the collision and had no hesitation. It seems that it was at least dark enough for him not to notice a person + a bike on a large road.
I'm also very (sadly) surprised that she crossed that kind of road at night without hurrying or reacting to the sound of cars approaching.
I too think visibility is better than it appears in the video, but I'm not so sure it's good enough to help all that much. However, even with visibility as bad as in the video, I'm confident in my ability to handle the situation. I would probably not be able to break in that short amount of time and from that speed, but neither would I drive at that speed. When there are less than ideal conditions (in this case visibility), it is our responsibility as drivers to adapt and lower the speed, possibly dramatically. This goes for autonomous cars too. If the road in front of me and the areas next to it are not clearly visible, I'd drive at such a speed that a collision would in all likelihood only result in scrapes.
True but there are trails that cross over the road, it is an odd area. If you zoom out on google maps you will see some of the trails. Note the sidewalk/pathway. It is no pedestrian but has paths for them so it sends mixed signals.
I saw that. The median is landscaped with a bizarre X-shaped paved area. It can't be intended for recreational use or walking; it's a divider between two fast roads. At all four entrances to the X, there is the no pedestrians sign.
I think you are wrong since this people die of this exact scenario almost everyday. The camera might be making it darker but that doesn't mean that every driver (Everyone's eyes and reaction times are different) would have been able to see her and get out of the way.
Is it possible that driving under intermittent street lights messes with the aperture or image recognition? It would be like flashing a strobe light at the camera.
right, so the camera's night vision mode that detects objects in the dark would have been completely blinded by the street lights while passing under the streetlamp. Take night vision goggles and look at a light. It blinds the whole field of vision.
The only thing that I think was the cars fault was that the car is programmed to drive when the driver is driving around distracted. There is no point to a human driver sitting behind the wheel of an autonomous vehicle if they aren't paying attention.
People need to understand that self driving mode isn't a freedom from the responsibility of driving safely. Rather its a tool to help ensure that driving statistically becomes safer as more self driving vehicles find their ways onto the road.
Hopefully someday all cars will be self driving and dangerous hazards/traffic reduced to the point that they are virtually none existent rather than being towards the top of the list of "preventable death" and "things humans don't want to waste most of their time during the day doing".
How did LIDAR and IR (?) not catch that? That seems like a pretty serious problem.
Something is badly wrong there. That should have been detected by LIDAR, radar, and vision. Yes, they need a wide dynamic range camera for night driving, but such things exist.[1][2] They're available as low-end dashcams; it's not expensive military night vision technology.
Radar should pick up a bicycle at that range. The old Eaton VORAD from about 2000 couldn't, but there's been progress since then.
LIDAR has its limitations; some materials, including the charcoal black fabric used on some desk chairs, are almost nonreflective to LIDAR. But blue jeans, red bike, bare head? Expect solid returns from all of those.
The video shows no indication of braking in advance of the collision. That's very bad. There simply is no excuse for this situation not being handled. The NTSB is looking into this, and they should. I hope the NTSB is able to pry detailed technical data out of Uber and explain exactly what happened. In the first Tesla fatal crash, they didn't get deeply into the software and hardware, because it was clear that the system was behaving as designed, unable to detect a solid tractor trailer crossing in front of the Tesla. The result of that investigation was that Tesla had to get serious about detecting driver inattention, like all the other carmakers with lane keeping and autobrake do.
This time it's a level 4 vehicle, which is supposed to be able to detect any road hazard.
The NTSB has the job of figuring out what went wrong, in detail, the way they do for air crashes.
LIDAR also has limitations on angular resolution just as a function of how the sensor works. It's entirely possible that the size of the person/bike on LIDAR was just too small until it was too late to stop.
Why it didn't even appear to try to stop? You got me, refresh rate on the LIDAR? LIDAR flat out being mounted to high and relying on optical sensors instead for collision avoidance of small targets (like a human head)?
I'm guessing, I'd love to see an NTSB report on this.
Why even bother having a LIDAR system on your self driving car if it doesn't have sufficient resolution to detect a person standing right in front of it?
This doesn't seem like an edge case at all. Pedestrian crossing the road at a normal walking pace, and no obstructions in the way which would block the car's vision. The fact that it's dark out should be irrelevant to every sensor on that car other than the cameras.
Something obviously went terribly wrong here; either with the sensors themselves or the software. Probably both.
For detecting larger obstacles like buildings or other vehicles would be my guess.
Realistically faster sensors should be used to detect obstacles. LIDARs I could find with some cursory googling can run up to 15hz. Computer vision systems can run much faster (I have a little JeVois camera that'll do eyeball tracking at 120hz onboard, I assume something that costs more can do better).
But more importantly, you're vastly trivializing the problem - Standing right in front of it, sure the LIDAR will see the person no problem. Standing 110 feet away (which would be min stopping distance at that speed)? Realizing that, for a LIDAR with a 400' range at 15hz moving at 40mph you get ~7 samples of a point before you're at it... For at least the first 3 frames that person is going to look like sensor noise. At 110 feet that person (which I'm calling a 2' wide target) is 1 degree of your sensor measurement.
It's not that it's useless or broken, more just this a seriously bad case where optical tracking couldn't work and where LIDAR is particularly ineffective at seeing the person because of how it works. More effective might be dedicated time of flight sensors in the front bumpers, unsure how long a range those can get, but they are also relatively "slow" sensors.
It’s not mutually exclusive either. You can have lower frequency, lower angular res 360 spinning LIDAR for low granularity general perception, and also have much higher frequency, brighter, and lower FOV (~90-120deg) solid state lidar mounted at the very least on the front corners of the car. We should be absolutely littering these vehicles with sensors, there’s no reason to be conservative at this stage.
> LIDAR also has limitations on angular resolution just as a function of how the sensor works. It's entirely possible that the size of the person/bike on LIDAR was just too small until it was too late to stop.
I highly doubt this is the issue. I am not sure what Ubers setup is, but even a standard velodyne should have been able to pick that up based on angular resolution.
> Realizing that, for a LIDAR with a 400' range at 15hz moving at 40mph you get ~7 samples of a point before you're at it... For at least the first 3 frames that person is going to look like sensor noise. At 110 feet that person (which I'm calling a 2' wide target) is 1 degree of your sensor measurement.
This is based on the velodyne LIDAR specs I could find last night with some quick googling:
- 400' range
- .04 degree angular resolution
- 15hz max update rate
If you have more accurate real world experience with these sensors and can share more accurate performance characteristics I can update.
These calculations were done assuming a vehicle moving at 40 mph. The stopping distance at that speed is about 110ft. I computed the pixel size by assuming 1 measurement = 1 pixel giving me 9000 pixels per 360 degrees.
Thats the one LIDAR Uber seems to have matching pictures.
5Hz - 20Hz full round sampling rate, lets assume 15 Hz.
The resolution in the horizontal plane is dependent on rotational speed, so at 15 Hz it should be 0,26 degrees.
(0,35/20*15 = 0.26)
For the woman height the angular resolution is 0.4 degrees no matter the rotation speed.
Id est, she would have been atleast one pixel wide from 400 feet and about 2 pixels high and growing in size if we assume 2' wide.
(Not counting bike).
I really see no exuse for Uber messing this up that bad. The LIDAR can't have missed a potential "obstacle" when it got closer, even if the car wouldn't classify it as a human.
I was using Rev E because it's the data sheet I had handy. Mostly I was trying to point out that LIDAR is not some magic thing that always sees everything and there's limitations.
There's with your .26 angular resolution @ 15hz. (I just have a spreadsheet that spits all these out for me.)
These are NOT big targets, they could easily have been mistaken for noise and filtered out. All of the LIDAR data I've ever seen has been fairly noisy and did require filtering to get usable information from it. And given the number of frames they get maybe their filtering was just too aggressive.
Yes, I agree with you that we can't assume that the car could have noticed the woman from 120 meters from LIDAR data alone. Maybe with some kind of sensor fusion with IR-cameras.
But, as it got closer and what the computer though was noise was on about the same place a sane obstacle finder should have given a posetive match. Maybe at 30 - 40 m worst case?
At 142 feet the woman probably had (assuming she was 5.5'):
asind(5.5/142) = 2.21* => 2.21/0.4 = 5.5
So between 5 and 6 "scanlines" going from left to right over her.
Assuming she was 2' wide that's 0.8 degrees which would be 2 to 3 pixels in breadth according to your spread sheet.
That's between 10 and 18 pixels (voxels?) that stand out clearly from the flat road around it, exluding the bike.
If you wan't to get an idea of how LIDAR data looks Velodyne has free samples and a viewer for less resolution models.
It pretty hard to identify obstacles far off, but you will still see there is something there. It's especially easy to identify obstacles that are vertical.
As she got closer, she would eventually show up clearly on the LIDAR data. But since the car never slowed down or went left, it didn't notice her at all even at point blanc (or did see her but failed to do anything about it).
A buddy of mine has a lower end LIDAR on a robot, working with them on SLAM on it, trying to get a similar hardware set up locally over the summer. (I have weird hobbies)
Yeah, I'm willing to accept SOMETHING bad happened here, as I said I really just wanna dissuade people from the notion that LIDARs will see all obstacles all of the time. Not going to say the car acted perfectly and it was sensor failure, but definitely willing to say that the LIDAR probably COULD see her but not as well as people would assume.
Really, I think this was a case of the car over driving their effective sensor range, same as what happens when you're on a dark road and a deer runs into the middle of the road, you simply can't react fast enough by the time you realize the danger is there. Computers are fast but they aren't perfect.
What I'd be particularly interested in was if the computer saw her and if it did the calculation - I can't stop safely in this distance, and decided to just hit the obstacle because it was "safer". At that point we start getting into ethics and this problem gets a lot murkier.
The person in that last picture is something like 5 feet from the car which is far to close to be useful at 40MPH. At those speeds what's important is what it sees at 150 and 200 feet and how fast it can refresh.
When your resolution is low enough to not see this they stop calling it LIDAR and start calling it a rangefinder. If this was actually a fundamental limitation of the sensor then that's the crappiest LIDAR unit I've ever heard of. When I first heard about the accident my initial reaction was "this is why vision only systems are inadequate". The fact that it didn't even detect the object at all before it collided with it with lidar, radar, and vision is inexcusable. This could set fully autonomous cars back enough that forget the cost of one life, this delay could kill tens of thousands because of a preventable accident.
I am glad to hear the NTSB is investigating this and not just the local police (who would lack the technical resources to make a useful judgement). Have there been any statements at all from them so far?
The NTSB tends to not make even initial statements until upwards of a several weeks in, and final statements as much as 6 months to a year later. They're nothing if not thorough.
Probably the only explanation was that LIDAR was off, either on purpose (for testing?) or because it broke and there was no safety mechanism to prevent the car from operating if/when LIDAR is off.
I don't understand why everyone in this thread is so focused on the sensors alone. The sensors might detect anything they like, but they're not going to stop the car on their own. The car has logic that tells it how to react to what its sensors perceive- that's the AI part. If the car's AI can't identify a woman pushing a bike as a woman pushing a bike, or it doesn't know that it has to stop before hitting her- well, then it won't.
There's so much confusion here, about the capabilities of these systems. People think that a combination of better sensory perception + faster reaction times suffices to drive in the chaos of the real world. That's not so. Sensors and fast thinking won't get you nothing if you can't think right. You have to be able to know what the things are that your sensors detect, and how to react to them.
It's perfectly possible that the Uber' car's LIDAR detected the lady crossing the road- but the AI just didn't know what to do about her and simply did nothing.
It's clear from the video that a human driver actually would've had more trouble since the pedestrian showed up in the field of view right before the collision, yet that's in the visible spectrum.
When I argue for automated driving (as a casual observer), I tell people about exactly this sort of stuff (a computer can look in 20 places at the same time, a human can't. a computer can see in the dark, a human can't).
Yet this crash proves that all the equipment in the world didn't catch a very obvious obstruction.