Featured Tesla removed Radar, now Sonar too.

When radar chips became sparse, Tesla suddenly stopped using radar in their cars, citing the difficulty of merging the logic of video based logic with data from a radar feed. Over the course of a few months they were able to replace much of the auto-pilot functions with video based environmental awareness. Now they are dropping the ultrasound sonar sensors that are used to locate nearby obstructions, including pedestrians. It looks like they are citing cost savings.

I find this change to be odd since it comes at a time when most car manufacturers are improving their ability to detect and avoid accidents by using more sensors and AI.

The link below has a video that shows that there are some really big blind spots in the car's camera field of vision. Most alarming is that there are areas in front of the car that would not see a child playing.

Tesla saves an estimated $114 per car by removing USS

 

No, I don't have a link that says that ultrasonics sensors are used to detect pedestrians, specifically. I gave up trying to nail them down on specifics when Musk was tweeting about how superior his business model was, allowing changes in manufacturing from day to day without having to remain backward compatible.

But I've seen MY car detect people walking in front of and behind my car when exiting a parking space. I'm pretty sure that's done by ultrasonic sensors. My radar, on the other hand, looks for cross traffic approaching from a distance and warns me in plenty of time even when I'm parked between two pickup trucks, SUVs, etc.

Relying on visual cues is probably easier to handle when you have an AI that's scrubbing video captures, but it seems to suck big time when you have low contrast scenes or when detecting things that are stationary and uncommon. The AI can't reliably interpret scenes that were never presented to it, right?

The problem I see with visual feeds for object detection is that the sample rate is relatively slow when you have to wait long enough for differences to show up between two frames. Only then can you correct for jitter and vibration and then finally apply algorithms to determine the distance from the parallax. How quickly can an optical system do that depending only on cameras? Can it plot the path of a car approaching from the left at 60 miles per hour before it enters an intersection? Judging from some of the YouTube videos that I've watched, the technique Tesla uses is often uncertain about oncoming traffic when the car has to turn left to merge into traffic. By uncertain I mean that it waits an extraordinarily long time for traffic to clear.

 

I asked: Can it plot the path of a car approaching from the left at 60 miles per hour before it enters an intersection? Judging from some of the YouTube videos that I've watched, the technique Tesla uses is often uncertain about oncoming traffic when the car has to turn left to merge into traffic. By uncertain I mean that it waits an extraordinarily long time for traffic to clear.

Mike was nice enough to reply:
Yes, it is being over cautious, perhaps, like someone with a learners permit.

Do you know whether this over cautious approach due to inability to determine the speed of the oncoming car? I suspect that it needs the cars to be in a certain number of frames over a certain period of time before the deltas can be accurately calculated.

I find that accurately assessing the speed of a car that is approaching head on is one of the challenges that many people do poorly. Guess wrong and you are liable to be in an accident. Quickly determining the speed of an approaching object is one of the major advantages that radar and lidar have.

 

From speedcam anywhere web site, the requirements for it to work are pretty restrictive:

Speedcam Anywhere needs to be able to get a clear view of the whole vehicle passing across the screen. It needs to have a clear view of the wheels, and to get an accurate measurement, a clear view of the numberplate as the vehicle passes. Aim the camera at about 45 degrees across the road for best results. The phone must be held still for 2 seconds whilst the capture takes place. Do not pan the camera to follow the vehicle. Practice in Basic Mode before using more credits in Pro Mode.

Lower spec phone cameras, rain, bad light and being further away from the vehicle will reduce the likelihood of the capture succeeding.

After all that, it has to send the video to a server farm where the actual calculations take place. I could probably calculate the speed quicker just based on known dimensions and how long it takes to cross the field of view.

But this does not appear to work for approaching traffic, nor for situations where the car with the camera is moving.