Uber self-driving car death riddle: Was LIDAR blind spot to blame? The death of a pedestrian in Arizona by an Uber self-driving car may have been the result of a blind spot caused by the use of a single LIDAR sensor on the roof. In 2016, Uber decided to shift from using Ford Fusion cars to Volvo XC90s for its self-driving car program. When it did so, it made big changes to its sensor design …
If there is a blind spot it is right in front of the bonnet. The lidar should have still seen the cyclist from far out. There is no info if the bike was carbon framed, I suspect not. If it was metal it would have been visible on the radar for miles. Similarly, most vis sensors are not actually vis, they stretch into near IR so there should have been no issue with the pedestrian having dark clothing and the car being driven dangerously for a human driver (to fast to stop within zone covered by headlights).
The only issue I can see is Uber. No other.
Similarly, most vis sensors are not actually vis, they stretch into near IR so there should have been no issue with the pedestrian having dark clothing
===============================================================
Not really.
1. A lot of cameras have an IR filter to keep infrared from reaching the sensor. Since the focus point for IR is different than for visible light, getting a clear image is enhanced by blocking the out of focus IR,
2. A visible light sensor may pick up near IR, it will not pick up far IR - generally such sensors, used in missiles, need to be cooled significantly to function, often with liquid nitrogen. Far IR is from 8 to 100 microns, and the peak emissions from the human body are at 9.3 microns. Unless the car has IR floodlights, near IR is no better than visible light with no lights.
.... A lot of cameras have an IR filter to keep infrared from reaching the sensor. Since the focus point for IR is different than for visible light, getting a clear image is enhanced by blocking the out of focus IR,...
You are talking about a camera optimised for visible light. Of course that has filters cutting out extraneous light that it's not meant to respond to.
But why on earth should a sensor INTENDED to pick up IR have an IR filter? That makes no sense. I am pretty sure that the design engineers would have speciified technology wich was designed to do its job. And that its job was to detect obstacles around the car in varying lighting conditions - which includes pedestrians in a night-time street. That there was a failure there is obvious. But I don't think it was as simple as using a technology which was not designed for the task....
All these sensors, and yet humans seem able to do this kind of thing quite a lot of the time with just a pair of mark 1 eyeballs and a g-sensor in the seat of the pants... it seems we can do remarkably complicated things with only a handful of sensors.
Though on a more serious note: I have wondered what happens when a lidar system, spitting out a presumably rather bright light (as far as its sensors are concerned) meets another lidar system. Oncoming headlights incorrectly adjusted or left on main beam are bad enough to cope with for a human driver; I'd love to know what an electronic system thinks of them. The same comment applies whether visible, IR, radar or indeed any other active lighting system is used.
I have wondered what happens when a lidar system, spitting out a presumably rather bright light (as far as its sensors are concerned) meets another lidar system
I think this is done by using specific frequencies for each car.
IR Filters..... most web cams are sensitive to IR, and an astronomer buddy of mine attached one to a telescope for imaging, after removing the IR filter. Photons are precious in astronomy, and IR is welcome. So yeah, you'd use as wide a spectrum as possible, you wouldn't filter anything out.
Not really, no. The whole sensor package looks mounted to the front section on the roof. There isn't any roof in front of it to block anything. I would expect the whole "blind spot" to be on the order of "something shorter than a foot, within two or three feet of the front bumper".
LIDAR maker CEO waxing about the indisputable need for more of its wares... laughable. That single one should have been perfectly capable of detecting the pedestrian, and I'm pretty sure it did too - at least at LIDAR reflection level. Why that never resulted in the car breaking is the actual million dollar question.
Even a carbon framed bike carries a large amount of metal.
Lidar <> radio frequency radar, so the metal or carbon fiber content are unlikely to be relevant. The question is: did the bike and pedestrian reflect laser light?
The question is: did the bike and pedestrian reflect laser light?
My question is more "what the fuck was the supervising meat-sack doing whilst this 'testing' was going on?". I'd have though that in any such test the human in the vehicle is still ultimately in charge else why be there at all? Not sure whether it was media bias and selective edits or not but the video I saw of the inside of the vehicle showed them paying zero attention to the road ahead when the accident occurred. Not really how a supervised test is really supposed to work.
Radar should have seen the pedestrian and bike.
Lidar would have picked up the bike and woman faster and easier than the radar.
Even still the car should have picked up the cyclist well within range of the sensor on the top of the car.
Drunk crawling. Victim of previous accident. Someone who tripped. Recumbent bicycle. Kid sitting on a skateboard.
Speed bump. Log in the road. Bricks, blocks, other obstructions.
An autonomous car has to scan for any and all obstructions ahead, and also be aware of hazards approaching from the sides and rear.
There is no excuse for not detecting someone wheeling a bike. Far smaller targets should be detected and acted on.
"There is no excuse for not detecting someone wheeling a bike. Far smaller targets should be detected and acted on."
And that's the bit that's really, really hard. Is it a paper bag? Is it a brick? Is it a plastic or glass bottle? Is it something that fell off the back of a lorry that might cause damage or and old newspaper just blowing in the wind?
"And that's the bit that's really, really hard. Is it a paper bag? Is it a brick? Is it a plastic or glass bottle? Is it something that fell off the back of a lorry that might cause damage or and old newspaper just blowing in the wind"
...and in the weird and wonderful real world of coincidences, I went out in the car a few hours after posting that and while driving had to avoid a claw hammer laying the road right where the wheels would normally be passing over. I've never seen one of those in the road before and it could have done serious damage either to the tyres or, if hit "right", the underside of the car, eg brake lines. Would an AV have seen it, recognised it for the danger it posed and avoided it?
Low is a relative term. Keep in mind a xc90 is 1766 mm or a couple inches shy of 6 feet tall and the LIDAR is roof mounted so it's likely easily over 6 feet. Depending on the vertical scan angle of the LIDAR there will be a blind spot created by a superposition of the roof of the car creating a largely rectangular cone the size of which is determined by how high the LIDAR unit is mounted and by the lowest angle it can scan.
If we assume the LIDAR is at 2 meters high and is mounted 1 meter from the edge of the roof (as the car is ~2 meters wide and I'm modelling it as a brick) this gives a maximum angle of 12.6 degrees below horizontal before the LIDAR beam is blocked by the roof outline. That puts the closest point on the ground the LIDAR unit can see at 8.9 meters (29.3 ft) away directly ahead and to the sides and 12.6 meters (41.4 ft) over the corner. If we move the LIDAR to 2.5 meters high these numbers drop to 3.5 meters (11.3 ft) and 4.9 meters (16 ft) respectively. The distance will be considerably greater toward the rear as the assumption here is that it is mounted 1 meter from the front edge of the roof and that would potentially place the rear edge considerably farther away decreasing the downward angle visible.
Yeah, this. Looking at the car it's mounted well above the roof line and should have a better view than a human driver. Assuming it's a 360 degree sensor as the story says it shouldn't have any issues picking out a pedestrian.
Indeed... People are taller than cars.
Even push bikes come up a good few feet, and from the footage, on that nice clear, uncluttered, and almost straight road, the entire bicycle and person were completely visible to the video camera.
To have a blind spot extend that far, the LIDAR would have to be at the rear of the roof, and as low down onto it as possible, which would be the stupidest location ever devised for a vehicle that will spend 99.99% of its time going forwards.
... and for some reason has never had to learn about consequences?
The only thing that makes me sad about Uber's ongoing fall from grace is that Kalanick and his bros have probably already cashed out, and they don't seem to be the sort of people to be overly concerned by the blood now on their hands.
diverting some of the investment into bank accounts as a share of the company is acquired is within the ability of most of these people.
Investors, yes but not employees who receive shares in lieu: their shares are untouchable and untradeable. Yet another way the VCs manage to shaft people. Didn't Kieren cover this a while back?
Hedgehogs. Why those haven't yet evolved titanium skin with tungsten spikes is one of Evolution's Great Mysteries.
A hedgehog is strolling the edge of a road, and comes across two rabbits. "Hey guys" he speaks up, "What is it that I see so much of my fellows being flattened while trying to cross, and only rarely a rabbit or a hare?". "Geometry", one or the rabbits answers. "If you're trying to cross at night, and you see headlights approaching, then either move back to the verge, or sit right in the middle of those lights. And by day you can use those headlights as markers too, even when they're off. Let me show you.". He jumps into the road, positions himself (looking over his outstretched front legs towards the approaching car's headlights), and indeed, the car passes right over him and he hops back to his mate and the hedgehog. "Aha" says the hedgehog, "Thanks, I'll try that". And as he starts crossing the road, another car approaches. The hedgehog shuffles around a bit, positioning himself and finds the right spot. "Oh look", the other rabbit remarks, "you don't see those Reliants much any more".
"Kittens!!!! You forgot kittens!!! Thank deity this isn't FB or you'd be toast."
For a more pragmatic approach (and a significant statistic): Toddlers playing on driveways
Not a problem at the moment, but it will be when robocars are the norm.
The lidar in question should easily have seen a person walking a bicycle in the middle of the road. There was some failure that wasn't caused by the lidar manufacturer's "not enough lidars" excuse. Maybe a dog right next to the car might be obscured by a fender, but not a five foot tall person walking a five foot long bicycle on a flat road sixty feet away.
There is also confusion about sensors being turned off. There are two sets of sensors: Uber's self drive sensors which are being tested and Volvo's proximity and braking sensors which are not. You can't test Uber's system if the Volvo system constantly interferes, so the Volvo system is shut off. The supplier of a Volvo system component is just pointing this out, so that their product does not get tarnished by Uber's mistakes.
You can't test Uber's system if the Volvo system constantly interferes, so the Volvo system is shut off.
This sounds like Uber is pretty bad at coding and debugging for real life situation.
In a testing lab, they could flip on their system and turn off Volvo system, where the worse case is you witness the car crash into stuff. However on real roads, it should be an OR case where if either system triggers the car brake, it brakes and a debug log is kept to verify if Uber's braking system was working.
It may be harder than that to setup an OR system. Volvo’s XC90 has an partial autonomous driving system of it’s own, not just automated emergency brakes. It steers the car (in lane), recognises speed signs, maintains following distance, etc. Separating out just the one part could cause bugs of it’s own too....
"The lidar in question should easily have seen a person walking a bicycle in the middle of the road. There was some failure that wasn't caused by the lidar manufacturer's "not enough lidars" excuse. "
You are assuming that the approaching human in question was in the 'cone of surveillance' for the lidar. With the reduction in sensors, that zone may have been relegated to another type of sensor, leaving the lidar for long range detection in the vehicle's planned path.
In that case the human would be invisible to the lidar until stepping in front of the car. Take a few fractions of a second to analyze and recognize, and you've run out of time.
You are obviously unfamiliar with the Velodyne unit used by Uber. Here is a somewhat dated link that will still help you visualize what that unit should have seen. Note that shadows appear behind large objects and very close to the vehicle, but that coverage is 360deg from a couple of meters to beyond the area depicted. The concentric rings are from the Velodyne.
https://www.youtube.com/watch?v=nXlqv_k4P8Q
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
