As a human drive I was fooled once by exactly the same things by exactly the same stretch of road, and almost hit that divider myself. It was one of the scariest moments of my driving ever. It's not the only time I misinterpreted those markings, too. I grew up with lanes marked by the pavement cracks. This trend of painting in lanes is new, and it's harder to perceive as a driver.
The problem with many of the freeways in the Bay Area, especially the Bayshore Freeway, is that the lanes have been moved many times since the original pavement was laid down, so seams and joints no longer align with painted markings. Yes, interpreting the lane positions can be a challenge, even for a human brain.
Dan Sanger The issue is not the state of the crash attenuator, the issue is the car drove itself into a stationary permanent and clearly marked median on a bright and sunny day detecting it 150 metres out and 6 seconds before the crash and inspite of warning the driver it could not avoid the coliision
This is especially bad when the sun is glaring on the road in morning/evening. 580 near Dublin has the same issue, and I did the same thing driving myself, causing me to drift into another open lane. The problem is the sun hitting the road at an angle at certain times of day makes the lane markings extra hard to make out, and that line/split in the road becomes much more distinguishable compared with the lane markings.
Good catch! Which is, again, why L2+ cars can not operate with CMOS cameras only but require a mix of cameras, radar, lidar and whatnot to be reasonably safe. "Autopilot" beats the texting-while-driving teen in accident statistics but unfortunately not (yet) the alert, responsible human...
I remember reading something about a year ago that a Volvo exec from Europe was astounded by how poor many of the road markings on US roads are. From what I understand, there was a nasty bus crash in that same spot and the human driving the bus couldn't see the stripes due to the rain. I do wonder if maybe CalTrans needs to not only restripe that section of road, but should cross stripe it and put arrester barrels filled with salt water there too until this gets straightened. out.
Whatever. Emergency braking should still apply and bring the Tesla to a stop in front of the clearly recognizable obstacle. Even if it misunderstands the road markings, instead of only alerting the driver to take over the steering, it should - well it MUST - apply the brakes in this situation to prevent this crash in the best manner possible. The obstacle is literally rock solid, it is clearly recognizable via the front radar. No sun glare on earth can explain this. I mean, that's what radar is made for: 3D imaging when normal cameras are in trouble. And for line detection: The lane detection cannot only use a Sobel filter to detect lines of whatever form. It must find the white or yellow lines of a reasonable width complying with the road standard in that particular country. IMHO: this video can in no way explain the crash.
Hi, can I show a clip from your video in a machine learning class I am teaching internal to my company (TD)? I will ink back here for students to see the full video.
Radar cannot detect fixed objects, in general. Radar only measures the difference in velocity between the car and the "thing" that is reflecting. For those "things," the radar can estimate the distance away, too. There are "things" all over the place that are reflecting back radar waves, and there's no way to tell them apart from the background (e.g., roadway). So basically radar removes anything that has a velocity that is the same as the background and looks at things that are moving either away from or toward the car (relative to the background). This is why radar-based TACC systems can pick up stopped cars that weren't previously being tracked (because they were seen moving before). It is also why radar-based cars will happily drive into another stopped car in your driveway or a wall or a median or a barrier. Detecting features that are obstacles requires vision (cameras and/or LIDAR), and the vision needs to be VERY confident that it is really seeing something before taking severe action to stop or veer. Vision is hard. Vision is used in Tesla AP in only very crude manner today; it only recognizes certain things like car/truck rear-ends, etc. They most likely then try to correlate that with radar image to confirm that a true obstacle is present, even if it isn't moving. That's why AP has recently started getting better at detecting stopped, previously-untracked cars (about 95% reliable, in my experience). Tesla's radar processing is extremely good, given the limits of the technology, but it is nowhere near able to detect obstacles, in general.
Too busy bashing Trump to gather your own facts? You fake journalists are a mess. Begging private youtubers for stuff, unbelievable. It is YOUR job to inform the public, get in a car, head there and do your job!
Comparing to the misread of lane markers, I'm more concerned that the software couldn't read the incoming obvious obstacle, painted in YELLOW AND BLACK. And there's a radar for that purpose in additional to cameras. This is the fatal mistake, I think it's an essential requirement that the software can detect a rock in the middle of lane, or a fallen furniture, before I can trust it with driving.
On the video, my car is on Autopilot. I drive the same section for nearly two years, 99.9% of time I'm on autopilot, however, this kind of error only happened to me once or twice. It's scary enough for me to keep high alert on this intersection.
All it takes is money, but the Democrats that control all branches of government in California have diverted highway maintenance funding that largely comes out of fuel taxes into other projects, like the High Speed Train to Nowhere and social programs for illegal aliens.
Milosz Ostrow So you want the state to use tax dollars to run around with a bucket of paint because a single player in a nascent technology builds a car with vision problems... something to benefit a tiny % of the driving population, rather than regulate that manufacturer to build a better car Get real.... that sounds like corporate (in the singular) welfare to me
Good to know this. What if the driver was on the left most lane? He works at Apple so probably on the left lane. Even if he was on the same lane as you are and autopilot got confused, then he was probaly inbetween lanes for sometime before autopilot alerted him. That mean he has more thna 5 seconds to react.
I think he was in the left lane - and i think the autopilot followed concrete/asphalt line as the video suggests, but in the left lane, the deviation from the white line is actually much much closer to the gore point, meaning he had much less time than 5 seconds.
Cody Trigg - Agreed. Concrete/asphalt line is probably providing the most contrasting and consistent edge. Most of the painted lines are inconsistent early on. Worse, the yellow line seems to disappear momentarily about the time the seam gets close to it. About this time, the right line provides a more convincing delineation of the right side of the "lane" and on up into the barrier. For a place that doesn't have to worry about snow plowing, frost heaving, and everything other than traffic destroying the roads, you'd think they could do a better job of painting lines (including cross hatching in between diverging lines) and maybe set out some crash barrels as a temporary fix until they get a replacement crush barrier.
This looks plausible. We don't really have much information at the moment. Many factors may be involved. "Auto-pilot" if it was on might have played a role but they ask you to keep your hands on the wheel for a reason. We don't know the speed and we also don't know if the driven might have been trying to avoid an obstacle on the lane (like a car dangerously changing lanes) it's not like you can grind to an halt on the highway. We don't know if they were under influence of some substance or if they suffered from a medical condition. It could have been intentional. Maybe the driver, for whatever reason wanted or was compelled in one way or another to veer left of the lane. Or maybe it was a mechanical failure. A tire bursting would cause asymmetric torque. It could be so many things, I wouldn't jump to conclusions.
Something doesn't add up. There are tons of youtube videos about autopilot failing, multiple times on the same trip. Imagine complaining to Tesla people "Well, it failed here and here and here and here." Do people really keep track of where autopilot failed? Most complaints I heard are general. In his case, the complaint was about this specific barrier? Weird to me. The best response he could get would be "well, we're working to improve it. Stay alert and keep your hands on the wheel."
As a human drive I was fooled once by exactly the same things by exactly the same stretch of road, and almost hit that divider myself. It was one of the scariest moments of my driving ever. It's not the only time I misinterpreted those markings, too.
I grew up with lanes marked by the pavement cracks. This trend of painting in lanes is new, and it's harder to perceive as a driver.
me too
The problem with many of the freeways in the Bay Area, especially the Bayshore Freeway, is that the lanes have been moved many times since the original pavement was laid down, so seams and joints no longer align with painted markings. Yes, interpreting the lane positions can be a challenge, even for a human brain.
and the crush barrier is still not fixed!
Great footage and explanation. Going to use this for my live show today, will give you credit!
There's really minimal protection on the end of that barrier. In my state, they usually install several plastic barrels filled with water or sand.
This is after the crash, I think that before it looks different.
Actually, there was a crash a week before and they didn’t have time to repair the softened crash area before the barrier.
no it's not. it is in fully compressed state. zero absorbation.
There was a crash there 10 days prior and the crash barrier hadn't been fixed
Dan Sanger
The issue is not the state of the crash attenuator, the issue is the car drove itself into a stationary permanent and clearly marked median on a bright and sunny day detecting it 150 metres out and 6 seconds before the crash and inspite of warning the driver it could not avoid the coliision
I get why the Autopilot would follow these lines, but why wouldn't it brake before hitting the barrier ? Can't it detect it ?
Yeah. This is another evidence that LIDAR is crucial.
Caltrans needs to fix the crash barrier and repaint the lines very clear.
A great explanation here...looks like an infrastructure problem to an extent
Great video. Still, why the autopilot doesn't see an obstacle on the (fake) lane and hit the brakes?
Exactly. If there was a LIDAR on the car, it might have been able to detect an obstacle.
I totally understand but why was the whole front of the Tesla gone it looked like it was going over 100 I auto pilot does not go ever 100
This is especially bad when the sun is glaring on the road in morning/evening. 580 near Dublin has the same issue, and I did the same thing driving myself, causing me to drift into another open lane. The problem is the sun hitting the road at an angle at certain times of day makes the lane markings extra hard to make out, and that line/split in the road becomes much more distinguishable compared with the lane markings.
Good catch! Which is, again, why L2+ cars can not operate with CMOS cameras only but require a mix of cameras, radar, lidar and whatnot to be reasonably safe. "Autopilot" beats the texting-while-driving teen in accident statistics but unfortunately not (yet) the alert, responsible human...
I remember reading something about a year ago that a Volvo exec from Europe was astounded by how poor many of the road markings on US roads are. From what I understand, there was a nasty bus crash in that same spot and the human driving the bus couldn't see the stripes due to the rain. I do wonder if maybe CalTrans needs to not only restripe that section of road, but should cross stripe it and put arrester barrels filled with salt water there too until this gets straightened. out.
Whatever. Emergency braking should still apply and bring the Tesla to a stop in front of the clearly recognizable obstacle. Even if it misunderstands the road markings, instead of only alerting the driver to take over the steering, it should - well it MUST - apply the brakes in this situation to prevent this crash in the best manner possible.
The obstacle is literally rock solid, it is clearly recognizable via the front radar. No sun glare on earth can explain this. I mean, that's what radar is made for: 3D imaging when normal cameras are in trouble.
And for line detection: The lane detection cannot only use a Sobel filter to detect lines of whatever form. It must find the white or yellow lines of a reasonable width complying with the road standard in that particular country.
IMHO: this video can in no way explain the crash.
Hi, can I show a clip from your video in a machine learning class I am teaching internal to my company (TD)? I will ink back here for students to see the full video.
Sure, no problem
Nice explanation. How do people fix this?
So basically CalTrans needs to paint fucking lines on the road. Got it.
Wouldn't the radar pick the median up?
Radar cannot detect fixed objects, in general. Radar only measures the difference in velocity between the car and the "thing" that is reflecting. For those "things," the radar can estimate the distance away, too. There are "things" all over the place that are reflecting back radar waves, and there's no way to tell them apart from the background (e.g., roadway). So basically radar removes anything that has a velocity that is the same as the background and looks at things that are moving either away from or toward the car (relative to the background). This is why radar-based TACC systems can pick up stopped cars that weren't previously being tracked (because they were seen moving before). It is also why radar-based cars will happily drive into another stopped car in your driveway or a wall or a median or a barrier. Detecting features that are obstacles requires vision (cameras and/or LIDAR), and the vision needs to be VERY confident that it is really seeing something before taking severe action to stop or veer. Vision is hard. Vision is used in Tesla AP in only very crude manner today; it only recognizes certain things like car/truck rear-ends, etc. They most likely then try to correlate that with radar image to confirm that a true obstacle is present, even if it isn't moving. That's why AP has recently started getting better at detecting stopped, previously-untracked cars (about 95% reliable, in my experience). Tesla's radar processing is extremely good, given the limits of the technology, but it is nowhere near able to detect obstacles, in general.
Thanks for the explanation,
But why do you have video footage from 8 days before the crash?
Good catch! Those pavement marks have probably messed up the Autopilot.
Are you cool if Channel 7 uses this video on the news tonight?
mightypowner
Can you call me? 415-672-9157
It will cost you
It's not yours.
dan noyes I don't make the rules Dan
Too busy bashing Trump to gather your own facts? You fake journalists are a mess. Begging private youtubers for stuff, unbelievable. It is YOUR job to inform the public, get in a car, head there and do your job!
Comparing to the misread of lane markers, I'm more concerned that the software couldn't read the incoming obvious obstacle, painted in YELLOW AND BLACK. And there's a radar for that purpose in additional to cameras. This is the fatal mistake, I think it's an essential requirement that the software can detect a rock in the middle of lane, or a fallen furniture, before I can trust it with driving.
Can you try with AP on to see how the car behaves?
On the video, my car is on Autopilot. I drive the same section for nearly two years, 99.9% of time I'm on autopilot, however, this kind of error only happened to me once or twice. It's scary enough for me to keep high alert on this intersection.
This is really a bad design, the space between the lines should be shaded.
Exactly, I just wonder whos' responsibility for this?
drive around in california. Most dividers aren't shaded in the middle. This isn't Europe.
I did. Didn't notice that TBH.
That this isn't Europe isn't the best argument ever.
All it takes is money, but the Democrats that control all branches of government in California have diverted highway maintenance funding that largely comes out of fuel taxes into other projects, like the High Speed Train to Nowhere and social programs for illegal aliens.
Milosz Ostrow
So you want the state to use tax dollars to run around with a bucket of paint because a single player in a nascent technology builds a car with vision problems... something to benefit a tiny % of the driving population, rather than regulate that manufacturer to build a better car
Get real.... that sounds like corporate (in the singular) welfare to me
Good to know this. What if the driver was on the left most lane? He works at Apple so probably on the left lane.
Even if he was on the same lane as you are and autopilot got confused, then he was probaly inbetween lanes for sometime before autopilot alerted him. That mean he has more thna 5 seconds to react.
I think he was in the left lane - and i think the autopilot followed concrete/asphalt line as the video suggests, but in the left lane, the deviation from the white line is actually much much closer to the gore point, meaning he had much less time than 5 seconds.
Cody Trigg - Agreed. Concrete/asphalt line is probably providing the most contrasting and consistent edge. Most of the painted lines are inconsistent early on. Worse, the yellow line seems to disappear momentarily about the time the seam gets close to it. About this time, the right line provides a more convincing delineation of the right side of the "lane" and on up into the barrier.
For a place that doesn't have to worry about snow plowing, frost heaving, and everything other than traffic destroying the roads, you'd think they could do a better job of painting lines (including cross hatching in between diverging lines) and maybe set out some crash barrels as a temporary fix until they get a replacement crush barrier.
This looks plausible. We don't really have much information at the moment. Many factors may be involved.
"Auto-pilot" if it was on might have played a role but they ask you to keep your hands on the wheel for a reason.
We don't know the speed and we also don't know if the driven might have been trying to avoid an obstacle on the lane (like a car dangerously changing lanes) it's not like you can grind to an halt on the highway.
We don't know if they were under influence of some substance or if they suffered from a medical condition.
It could have been intentional. Maybe the driver, for whatever reason wanted or was compelled in one way or another to veer left of the lane.
Or maybe it was a mechanical failure. A tire bursting would cause asymmetric torque.
It could be so many things, I wouldn't jump to conclusions.
Something doesn't add up. There are tons of youtube videos about autopilot failing, multiple times on the same trip. Imagine complaining to Tesla people "Well, it failed here and here and here and here." Do people really keep track of where autopilot failed? Most complaints I heard are general. In his case, the complaint was about this specific barrier? Weird to me. The best response he could get would be "well, we're working to improve it. Stay alert and keep your hands on the wheel."
That concrete divider looks like a bad design. It's simply a sudden stop. It should instead have some type of diversion system.
This is after the crash, I think that before it looked different.
There is a reset-able crash barrier but Caltrans is too lazy to replace it.
and another reason is the bumper was brocken than it's usual, as a car crash it on 12/3/2018.
If you’re not going to watch the road closely when using autopilot, you should at least be in the right or middle lane, following a lead car.
I've been saying the same things since the day of the crash.
I hate concrete highways so much. Hard to see even as a person.
Crash. Not crush.