Cyclist killed by self-driving car while walking her bike in AZ

[trailrunner]

I have done a little work developing sensors for autonomous vehicles. I am absolutely not surprised that this happened.

Sent from my iPhone using Tapatalk

[Brendan von Buckingham]

Paving parts of the median to look or act like pedestrian crossings, and then posting a sign that says “No pedestrians. Use crosswalk” (but providing no sidewalk to safely get you to the crosswalk) is terrible design. Or entrapment. Sue everybody.

[AFHokie]

https://www.sfchronicle.com/business/article/Exclusive-Tempe-police-chief-says-early-probe-12765481.php?utm_campaign=twitter-premium&utm_source=CMS%20Sharing%20Button&utm_medium=social

From the article link: The incident happened within perhaps 100 yards of a crosswalk, Moir said. “It is dangerous to cross roadways in the evening hour when well-illuminated, managed crosswalks are available,” she said.

Sent from my SAMSUNG-SM-G930AZ using Tapatalk

[mstone]

unlike most collisions, this one will be well documented. figuring out who screwed up shouldn’t be that hard, and with much less hearsay than usual.

[KWL]

@AFHokie 176548 wrote:

From the article link: The incident happened within perhaps 100 yards of a crosswalk, Moir said. “It is dangerous to cross roadways in the evening hour when well-illuminated, managed crosswalks are available,” she said.

A crosswalk 100 yards away (length of football field) is not one that is “available.”

[trailrunner]

So what if the pedestrian was or was not in a crosswalk?

Walking at night may or may not have been an issue. A common sensor for autonomous vehicles is a LIDAR, which is an active system which will work day or night. If the vehicle was using a passive sensor, many of the commercial systems in cars use an IR sensor, which should work at night.

The part that interests me is that she was walking her bicycle. Having that in the scene along with the human might’ve confused the algorithm.

However, there are many other factors that could have led to the failure. Image processing is complex, and dynamic urban scenes are extremely complex. Humans are very good at processing scenes (although not perfect, of course). Teaching a machine to do that, and to work in every possible situation, is extremely challenging.

[Brendan von Buckingham]

If she was following the paved sidewalk in the median and the desire trails worn into the landscape on the opposite side of the road she had probably dismounted her bike to cross the road. It probably wasn’t that she was just walking along the side of the road. I bet local human drivers know that people regularly cross at this unofficial crossing. Which is why authorities put up the no pedestrian sign. Unless the no pedestrian sign is one of those non-official (no certification stamp on the reverse) signs like the stop signs on the Custis trail put up by the Key Marriott at their entrance.

[DrP]

Camera footage now released. See on NPR page: https://www.npr.org/sections/thetwo-way/2018/03/21/595941015/police-in-arizona-release-dashcam-video-of-fatal-crash-involving-self-driving-ca

Looks like (1) she was crossing from the paved median to the side with the desire trails. (2) The driver was not paying attention at all. And likely (3) the sensors either weren’t working at all or the algorithm is not good enough. (as someone who works with/develops sensors and algorithms to find things in data, I know these things are only as good as all the cases the develops thought of, if that good. Pedestrians not at cross walks at night should have been thought of and designed for).

[Steve O]

@DrP 176760 wrote:

Looks like (1) she was crossing from the paved median to the side with the desire trails.

This is a design clusterf**k that was going to kill someone eventually, regardless of whether the car was self-driven or human-driven.

There’s a friggin’ pathway in the median that leads to this point! The highway design essentially ENCOURAGES people to cross here.

[ATTACH=CONFIG]17686[/ATTACH]

[mstone]

@DrP 176760 wrote:

Camera footage now released. See on NPR page: https://www.npr.org/sections/thetwo-way/2018/03/21/595941015/police-in-arizona-release-dashcam-video-of-fatal-crash-involving-self-driving-ca

Looks like (1) she was crossing from the paved median to the side with the desire trails. (2) The driver was not paying attention at all. And likely (3) the sensors either weren’t working at all or the algorithm is not good enough. (as someone who works with/develops sensors and algorithms to find things in data, I know these things are only as good as all the cases the develops thought of, if that good. Pedestrians not at cross walks at night should have been thought of and designed for).

Honestly, I have to fault the ped on this one. They are literally right at the edge of a lit segment, in the shadows between the lights–I’m pretty careful behind the wheel but I’m not sure I would have seen her either. https://www.youtube.com/watch?v=pO9iRUx5wmM Look at the time between 2 and 3 seconds–the feet basically appear out of nowhere. There are no reflectors on the bike wheels or tires. And the pedestrian doesn’t seem to be looking up at the lights on the car, or reacting at all–did they just assume the car would stop? Clearly the driver was being paid to do a job that they weren’t actually doing, but I’m not sure how much of a difference that made.

Just like drivers don’t understand how dangerous a lot of what they do is, I don’t think pedestrians/cyclists always understand how hard it is for drivers to see something that’s not lit or reflective when driving on a road filled with a lot of brightly lit objects that are making it hard to see darker objects–the pedestrian can see fine and assumes that the driver can as well, but that’s simply not the case. (That only gets worse when a young cyclist is using what they can see to guess what an older driver can see.) And street lights only make things worse if you’re not under the streetlight. I know a lot of people take the reflectors off their wheels because they’re uncool or something, but really the reflectors are usually the only thing I see when approaching a cyclist. (Dim blinkies are way less visible than a reflector bouncing back a significant fraction of the ~1400 lumens I’m sending down the road.) Wheel reflectors in particular are the only thing that will typically be visible when approaching from the side as in this case (even “side facing” battery lights are decorative rather than functional).

Also, the comment about the road design being fatally bad is absolutely correct. At the very least, there should be a light at the crossing. Except then they wouldn’t be able to pretend that there isn’t a crossing. I would predict that their response, if any, will be a fence.

[mstone]

@Brendan von Buckingham 176510 wrote:

Paving parts of the median to look or act like pedestrian crossings, and then posting a sign that says “No pedestrians. Use crosswalk” (but providing no sidewalk to safely get you to the crosswalk) is terrible design. Or entrapment. Sue everybody.

I can’t figure out WTF that big X on the median is supposed to be. My best guess is that maybe there was a crossover there to facilitate closing one side of the road at a time for bridge construction/maintenance and they kept it in case they ever needed it again. It’s the width of two traffic lanes, doesn’t connect to anything, and doesn’t seem to serve any current purpose. But there’s a streetlight in the middle of it.

[LhasaCM]

@DrP 176760 wrote:

Camera footage now released. See on NPR page: https://www.npr.org/sections/thetwo-way/2018/03/21/595941015/police-in-arizona-release-dashcam-video-of-fatal-crash-involving-self-driving-ca

Looks like (1) she was crossing from the paved median to the side with the desire trails. (2) The driver was not paying attention at all. And likely (3) the sensors either weren’t working at all or the algorithm is not good enough. (as someone who works with/develops sensors and algorithms to find things in data, I know these things are only as good as all the cases the develops thought of, if that good. Pedestrians not at cross walks at night should have been thought of and designed for).

Agreed. I think it was Wired that had an article about how this is precisely the type of collision that self-driving cars with sensors and everything are supposed to prevent. There was however long a “slow walk across at least 1 1/2 lanes” of traffic take for the computer to figure it out from the advanced sensors like LIDAR. Even from the video, it looked like there was about 2 seconds from when she came into view in the headlights and the impact; the notion that the car didn’t even try to slow down shows that even the basic processing of the situation had failed.

[DrP]

@mstone 176767 wrote:

Honestly, I have to fault the ped on this one. They are literally right at the edge of a lit segment, in the shadows between the lights–I’m pretty careful behind the wheel but I’m not sure I would have seen her either. https://www.youtube.com/watch?v=pO9iRUx5wmM Look at the time between 2 and 3 seconds–the feet basically appear out of nowhere. There are no reflectors on the bike wheels or tires. And the pedestrian doesn’t seem to be looking up at the lights on the car, or reacting at all–did they just assume the car would stop? Clearly the driver was being paid to do a job that they weren’t actually doing, but I’m not sure how much of a difference that made.

Just like drivers don’t understand how dangerous a lot of what they do is, I don’t think pedestrians/cyclists always understand how hard it is for drivers to see something that’s not lit or reflective when driving on a road filled with a lot of brightly lit objects that are making it hard to see darker objects–the pedestrian can see fine and assumes that the driver can as well, but that’s simply not the case. (And that only gets worse when a young cyclist is using what they can see to guess what an older driver can see.) I know a lot of people take the reflectors off their wheels because they’re uncool or something, but really the reflectors are usually the only thing I see when approaching a cyclist. (Dim blinkies are way less visible than a reflector bouncing back a significant fraction of the ~1400 lumens I’m sending down the road.) The wheel reflectors are the only thing that will typically be visible when approaching from the side as in this case. And the street lights only make things worse if you’re not under the streetlight.

Also, the comment about the road design being fatally bad is absolutely correct. At the very least, there should be a light at the crossing. Except then they wouldn’t be able to pretend that there isn’t a crossing. I would predict that their response, if any, will be a fence.

Possible ped was at fault or partially at fault – as a ped I always cross very quickly in situations where I do not have a walk signal (and sometimes when I do), but oh so many people seem to feel that they can stroll or saunter. I do not know the expectation there – some places have more expectation than others.
The images we see on youtube are not necessarily what the person in the drivers seat will see nor what the sensors can detect – I find that camera images are typically darker than what my eyeballs see. And the sensor system should have a greater ability to sense differences.

I do wonder what the vehicle would do around here with deer. Or anywhere with other animals. They do not wear reflectors (if I can figure out how to get the W&OD herd to wear them, I will try it). They appear to walk, jump or run suddenly into a path. They are difficult to detect, but if self driving cars continually hit deer, they won’t be used. If the algorithms and sensors could sense deer, this cyclist would probably not have been hit.

[mstone]

@DrP 176770 wrote:

Possible ped was at fault or partially at fault – as a ped I always cross very quickly in situations where I do not have a walk signal (and sometimes when I do), but oh so many people seem to feel that they can stroll or saunter. I do not know the expectation there – some places have more expectation than others.
The images we see on youtube are not necessarily what the person in the drivers seat will see nor what the sensors can detect – I find that camera images are typically darker than what my eyeballs see.[/quote]

yes and no–cameras typically have more limited range than human eyes, but human eyes vary widely from person to person.

Quote:

I do wonder what the vehicle would do around here with deer. Or anywhere with other animals. They do not wear reflectors (if I can figure out how to get the W&OD herd to wear them, I will try it). They appear to walk, jump or run suddenly into a path. They are difficult to detect, but if self driving cars continually hit deer, they won’t be used. If the algorithms and sensors could sense deer, this cyclist would probably not have been hit.

It’s kinda important to realize that the cars are still in development and not finished products. And uber’s program is nowhere near the most mature. (I honestly don’t know why they even have one, except that they have too much cash.)

[scoot]

For one thing, it is important to keep in mind that this video displays only a subset of the data that would have been available to the AV sensors and its human driver. It is trivial to play with video histograms or display settings so that they show or hide exactly what the person releasing a video wants them to. Even if nothing was altered intentionally, it is extremely unlikely that any display configuration chosen for presentation would manage to encompass the entire dynamic range of the scene that was actually perceptible to the sensors and/or to human vision.

But let’s ignore that argument for a second, and assume that the victim could in fact not have been detected at all before the instant she appears in the video. What this now means is that this vehicle is driving way too fast for its own headlights. The dashed lines are disappearing at about 90 Hz. When the victim’s feet are first perceptible in this video, she is a little over 1.5 dashed-line cycles ahead of the car. The car covers that distance in a little over one second and strikes her.

This pedestrian did not jump out from behind an obstacle; she had already crossed the super-wide area where the left lane opens up into two turn lanes. She had probably been in the northbound roadway for five seconds before impact.

Is it responsible to drive at a speed that allows you to see only one second in front of you? The recommended highway following distance is two seconds, and that’s when you’re behind someone who’s moving the same direction you are!

[Author]

Posts