Here is an analysis from an SAE discussion group, the vision system used would not pass a basic drivers eye exam.
I can try and download the example pictures if anyone is interested.
Quote:
My concern is the instrumentation is substantially inferior to a human. For instance, the forward looking camera in the Tesla is HD resolution (1280x720) with a wide angle lens that has a 5.6 arc-min resolution per pixel, compared to a human which has 1 arc-min. Or to put in human terms, the Tesla auto-pilot would fail the DMV vision test. I estimate the Tesla has 20/110 vision, or in other words it can see the big E at the top of the eye chart, but will make a couple mistakes on the second line.
In the Florida accident, the Tesla came over the crest of the hill just under 1/2 mile (2250') from the truck. If it were a human driving, when cresting the hill it would be obvious that a large tractor-trailer is crossing in front of them. The human would let off the gas, and give the truck the extra 5 seconds in needed to clear the intersection. But for the computer the truck is a 8x3 pixel blob at 2250'. At 1/4 mile, it's still an unidentifiable blob -- 15x4 pixels, something is there, but not enough data to identify it. At 1/8 mile, there's finally enough data (30x8, 240 pixels) to identify it's a truck in optimal conditions. It's not till 500ft, the computer can clearly identify it's a truck, but the writing/text on the trailer would still be blurry. At 250ft, only the largest letters (>4ft tall) on the trailer would be legible. Not until 100ft, is there any detail. Attached below is a simulation of what a tractor-trailer looks like.
|
When all vehicles are autonomous they called all be linked, sharing position and velocity data with every vehicle in the vicinity, they wouldn't have to see, but would know where every vehicle is, and where it will be, avoiding an event like this.