I believe that Autopilot with 5 cameras covering 360 degrees is possible, the only roadblock is figuring out the cost structure that will make this work. One high frame rate camera in front, 4 lower frame rate camera in each corners of the car. Keep cost below $5000 and use image processing to determine distance instead of a Lidar. I don't claim to be an expert in this field, but I've done some work in image processing with the 360 degree camera system that Elon mentioned. Hopefully someone better than me can find some fault and we can debate about possibilities.
Hardware setup
If I remember correctly, the two major players in image processing hardware space are Cognex and Matrox. Cognex has similar offerings at a lower cost, but Matrox has a better API.
•One capture card with 8 input feeds ~$2000
•One high frame rate camera ~$1000
•Four low frame rate cameras ~$2000
•Bumper sensors piggy backing on airbag sensors ~$0
•Shared CPU with car console ~$0
Five cameras with one high frame rate in front and 4 lower frame rate cameras at the four corners of the car with overlapping field of view. Autopilot can piggyback on the sensors used by airbags and use them as a collision detector for last line of defense, saving some cost. The biggest variable in the cost, is in how much camera we need based on the requirement. Obviously, the more camera we slap on, the cheaper the camera and image quality if we want to keep the cost to $5000.
The front high frame rate camera's main job is to determine relative speed and prevent a frontal collision, this is where things gets a bit fuzzy, we need to figure out what frame rate is necessary for this to function properly and what transfer delay from capture to the cpu's memory is acceptable. For a surgeon's medical console, I know that the system needs to be 60hz, 10bit color and 30 msec maximum delay between capture to display. So we can start with that as a reference design.
Roadblock 1: Relative distance of moving cars
The next problem is sensing distance. This is the problem that Lidar solves and a camera based system might have problems with. With cameras, it is hard to determine how close an object is, because an object can just be big and cover a large pixel area of the sensor without being near the car. The one way I can think of is to use image processing techniques to process this information. Here is the trick I think will solve the problem of how to determine relative distances of cars nearby: License plates. License plates are state issued and have a standard size. It is also something that is on every car and has a white background which is very easy to be picked out by image processing software.
Roadblock 2: Pedestrians
Pedestrians present a different challenge as they are mostly immobile when looked at from the point of view of the car. The most important thing about pedestrians is determining their distance to the car. Assuming they are immobile, we can use the difference of size between subsequent frames and the car's own speed to determine how fast the pedestrian is approaching the car. Again, each person is different and their height will affect the judgment of this process, but a person can probably be estimated based on how far off the vertical axis they are in the image sensor and how much vertical pixels they occupy.
Cost breakdown
I am only familiar with Matrox's cost, so using the Matrox QxT line of product where 8 inputs can be received and captured by one PCIe card connected to a PC, we should be able to reduce the cost of electronics, pushing the majority of the system's cost onto the hands of the cameras themselves. Here's the tricky part that I believe can make or break the bank.
If my memory serves me correctly, an Dalsa camera normally paired with these capture cards in an industrial setting can cost up to $10,000 with the lens included. Multiply that by 8 and we have the autopilot system that's equivalent to the cost of a Model S. As a consumer, it does not make sense to buy that... unless you are in Norway. From a back of the napkin calculation, Model S owners might be willing to shell out $5,000 for technology based on the evidence that most people choose the tech package that cost around $3,750. $5,000 on a $75,000 Model makes sense since that's only 6% of the overall cost. $5,000 on Gen 3 that cost $35,000 might not make sense for people as that represent a full 14% of the cost.
So let's limit the scope to model S for now. With Matrox's Qxt, the cost of each capture card is around $2,000. Assuming we are using the on board CPU to keep cost in line, we have $3,000 left over to buy 5 cameras and we are left with off-the-shelf type consumer type cameras. Personally I think it is fine, because the on board CPU can't handle the load of processing 8 high def cameras anyway.
Disclaimer
All opinion and information are for Tesla Motors Club discussion only. Please do not quote me anywhere else, nor use the information for any type of formal debate. I don't claim that I know anything, nor do I claim that any of the cost here are accurate. These are back of the napkin estimates and information pulled from my memory from something I did a couple of years ago.
