3 ms·

Your comment actually emphasizes why having a lidar is better. Cameras need neural networks to detect objects. Lidars on the other hand provide distance to all

Your comment actually emphasizes why having a lidar is better. Cameras need neural networks to detect objects. Lidars on the other hand provide distance to all the surrounding objects. The neural networks processing the point cloud may not be able to detect what the object is, but because of the distance available in the point cloud data you will know there is something ahead. Fusing radar with camera does help but radars are noisy and unreliable.

CardenB 7y ago
Lidar is also noisy unfortunately. The input is high dimensional enough that classical tracking methods don’t work well (like for simpler radar systems) because segmentation in lidar is hard.
bobsil1 7y ago
Lidar is active (so lower res than vision) and uses slightly longer wavelengths (again, lower res). Visible light uses passive solar illumination by day, power-efficient, far more data; and is a problem you need to solve for reliable autonomy independent of lidar.
- studentrob 7y ago
  > is a problem you need to solve for reliable autonomy independent of lidar Not for L4. A level 4 vehicle is not expected to operate under all conditions, including weather and location.
sjwright 7y ago
Radar seems to work pretty damn well for the adaptive cruise control and autonomous emergency braking installed in hundreds of models of cars made over the past 5+ years. If autonomous systems needs more redundancy, the solution is to keep adding more cameras. The other sensor which I think is important is data captured by cars in that fleet. This is sort of what's known as "HD maps" but fully automated and based purely sensor-derived data. So if you're driving down a highway at night in the rain, your car would know the successful paths traversed a few hours ago when it was dry and sunny.
- jacquesm 7y ago
  > autonomous emergency braking I've just gotten rid of a car that did that and that was one of the contributing factors. That system was so totally unreliable that it almost caused two accidents in 20K km when otherwise nothing would have happened. In both cases there were obstacles on the side of the road that gave a return (bridge supports/railing and a pretty sturdy billboard frame) to the radar that caused it to slam the brake full on. On the bridge that caused the car to skid , I narrowly avoided the posts on the far side of the bridge and the second instance the person behind me was totally surprised (as was I). That sort of tech is really not yet ready for deployment as far as I'm concerned, it's a liability instead of extra safety. I contacted the manufacturer and they saw nothing wrong with the car. The system could not be disabled either.
  - areoform 7y ago
    Sadly, not all ACC and CAS deployments are created equal. Tesla uses a 77ghz RADAR that has 0.1° of accuracy on the azimuth for the far range radar at a beam spread of 6° and 0.2° for 9°. The near RADAR resolves to between 0.3° to 5° for the close one. They can pick out objects at a resolution of 40cm for the long range RADAR and down to 5cm short range. http://www.compotrade.ru/i/pdf/ARS404-21_ARS408-21_en_V1.03.pdf http://www.compotrade.ru/i/pdf/ARS404-21_ARS408-21_en_V1.03.... Not all manufacturers use such high specced RADAR as you can see in this comparison sheet there are cheaper options to pick from that don’t offer anywhere close to this angular accuracy and resolution https://autonomoustuff.com/wp-content/uploads/2018/02/RADAR-Comp-Chart-2018.pdf https://autonomoustuff.com/wp-content/uploads/2018/02/RADAR-... The manufacturer is probably using 24ghz RADAR which is legacy technology and is being discontinued due to lower accuracy. There are software tweaks available to reduce false positives for legacy tech but most manufacturers don’t deploy them. See: https://e2e.ti.com/blogs_/b/behind_the_wheel/archive/2017/10/25/why-are-automotive-radar-systems-moving-from-24ghz-to-77ghz https://e2e.ti.com/blogs_/b/behind_the_wheel/archive/2017/10...
  - sjwright 7y ago
    There might be some bad implementations out there, but that is a problem with that implementation, not the technology. I've driven extensively with AEB in a 2015 Jeep Cherokee and it has been borderline faultless. And that's surprising because I have very low regard for anything made by Fiat/Chrysler, but whatever (presumably third party) tech they integrated, it has been rock solid. Over tens of thousands of kilometres it has never, ever had a true false positive. And the radar cruise has never, ever had a false negative. Very occasionally the AEB has "freaked out" unnecessarily; usually when I'm driving right into someone who is in the process of turning off the road. I know that they'll be out of my way in time for me to go past, so technically it's a false positive, but even then I'd give it to the car. I am driving dangerously. I am gambling that the turning driver doesn't brake suddenly before they complete their turn (I've also done a few thousand kilometers in a 2019 Subaru Forester with Eyesight—a purely vision-based system—and its adaptive cruise is even better than on the Jeep. It's less anxious on the brakes in response to movements by the cars ahead, and you can configure how eagerly it accelerates back up to set speed when the road clears.)

CardenB 7y ago

Lidar is also noisy unfortunately. The input is high dimensional enough that classical tracking methods don’t work well (like for simpler radar systems) because segmentation in lidar is hard.

bobsil1 7y ago

Lidar is active (so lower res than vision) and uses slightly longer wavelengths (again, lower res). Visible light uses passive solar illumination by day, power-efficient, far more data; and is a problem you need to solve for reliable autonomy independent of lidar.

studentrob 7y ago
> is a problem you need to solve for reliable autonomy independent of lidar Not for L4. A level 4 vehicle is not expected to operate under all conditions, including weather and location.

sjwright 7y ago

Radar seems to work pretty damn well for the adaptive cruise control and autonomous emergency braking installed in hundreds of models of cars made over the past 5+ years. If autonomous systems needs more redundancy, the solution is to keep adding more cameras. The other sensor which I think is important is data captured by cars in that fleet. This is sort of what's known as "HD maps" but fully automated and based purely sensor-derived data. So if you're driving down a highway at night in the rain, your car would know the successful paths traversed a few hours ago when it was dry and sunny.

jacquesm 7y ago
> autonomous emergency braking I've just gotten rid of a car that did that and that was one of the contributing factors. That system was so totally unreliable that it almost caused two accidents in 20K km when otherwise nothing would have happened. In both cases there were obstacles on the side of the road that gave a return (bridge supports/railing and a pretty sturdy billboard frame) to the radar that caused it to slam the brake full on. On the bridge that caused the car to skid , I narrowly avoided the posts on the far side of the bridge and the second instance the person behind me was totally surprised (as was I). That sort of tech is really not yet ready for deployment as far as I'm concerned, it's a liability instead of extra safety. I contacted the manufacturer and they saw nothing wrong with the car. The system could not be disabled either.
- areoform 7y ago
  Sadly, not all ACC and CAS deployments are created equal. Tesla uses a 77ghz RADAR that has 0.1° of accuracy on the azimuth for the far range radar at a beam spread of 6° and 0.2° for 9°. The near RADAR resolves to between 0.3° to 5° for the close one. They can pick out objects at a resolution of 40cm for the long range RADAR and down to 5cm short range. http://www.compotrade.ru/i/pdf/ARS404-21_ARS408-21_en_V1.03.pdf http://www.compotrade.ru/i/pdf/ARS404-21_ARS408-21_en_V1.03.... Not all manufacturers use such high specced RADAR as you can see in this comparison sheet there are cheaper options to pick from that don’t offer anywhere close to this angular accuracy and resolution https://autonomoustuff.com/wp-content/uploads/2018/02/RADAR-Comp-Chart-2018.pdf https://autonomoustuff.com/wp-content/uploads/2018/02/RADAR-... The manufacturer is probably using 24ghz RADAR which is legacy technology and is being discontinued due to lower accuracy. There are software tweaks available to reduce false positives for legacy tech but most manufacturers don’t deploy them. See: https://e2e.ti.com/blogs_/b/behind_the_wheel/archive/2017/10/25/why-are-automotive-radar-systems-moving-from-24ghz-to-77ghz https://e2e.ti.com/blogs_/b/behind_the_wheel/archive/2017/10...
- sjwright 7y ago
  There might be some bad implementations out there, but that is a problem with that implementation, not the technology. I've driven extensively with AEB in a 2015 Jeep Cherokee and it has been borderline faultless. And that's surprising because I have very low regard for anything made by Fiat/Chrysler, but whatever (presumably third party) tech they integrated, it has been rock solid. Over tens of thousands of kilometres it has never, ever had a true false positive. And the radar cruise has never, ever had a false negative. Very occasionally the AEB has "freaked out" unnecessarily; usually when I'm driving right into someone who is in the process of turning off the road. I know that they'll be out of my way in time for me to go past, so technically it's a false positive, but even then I'd give it to the car. I am driving dangerously. I am gambling that the turning driver doesn't brake suddenly before they complete their turn (I've also done a few thousand kilometers in a 2019 Subaru Forester with Eyesight—a purely vision-based system—and its adaptive cruise is even better than on the Jeep. It's less anxious on the brakes in response to movements by the cars ahead, and you can configure how eagerly it accelerates back up to set speed when the road clears.)