I think this is indeed the case. In my area there are 4 lane roads where the average speed is 60mph. For an unprotected right on red at a traffic light the cameras have to look across a wide intersection and then work out the velocity of cars that are approaching the traffic light - about 100m. At least with 10.12 it seems to not see the traffic at all - when it is 100% clear it pauses a good 15 seconds and does a timid commit to the turn. When there is approaching traffic it does the same thing - except there is traffic and I have to disengage or slam on the accelerator / change lanes. Curious to see if 10.69 is any better on this type of turn.
Phantom breaking is probably similar - at 1280x960 your are calculating the risk of hitting fuzzy pixelated blobs at 50 meters - AI's can do amazing things but it will be limited by the cameras. If the cameras were sharper there would be much better distinction of shadows / lighting vs a concrete object in the road at distance. And yeah, the compute on higher res cameras would be way too much.
Um. There’s this thing in traditional DSP audio processing where one is trying to recognize diphthongs and words and such.
As some of you who are older may remember, getting those early algorithms to work reliably involved training with the user with many sounds, and if users were changed, the user who had done the original training had a cold, or whatever, the accuracy fell off a cliff. And it wasn’t that great in the first place. Think early versions of Dragon Dictate.
One day, this all changed dramatically: the algorithms used began to use Noisy Markov Chains math.
A Markov chain algorithm has the idea that one knows the past history of a number of events, a probability matrix of what the
next event might be, and the current raw data input. Into such an algorithm one could place a series of already detected diphthongs, a probability matrix based upon the language being used, and the current raw detected audio. It worked, but was not much better than Dragon Dictate.
The breakthrough was to realize that one was looking at an ideal bunch of diphthongs that were masked with colored noise, the noise being generated by variations in peoples’ vocal tracts, the ambient, and so on. There’s some serious math involving detecting signals in the presence of noise (think: Information Theory, a la Shannon), which could then be applied. Suddenly, there was a quantum jump in the effectiveness of voice recognition technology, where speaker-independent voice recognition technology could be deployed-and it Just Worked. Giving rise to Iron Ladies everywhere and, eventually, for those minuscule processors on your cell phone to correctly respond to, “Hey Google!” or whatever.
Thing is, a low resolution camera is looking at a full resolution image. The image on the CCD is the real image with (wait for it..) quantization noise. Which is there because the real world is continuous, but the image sensor is discrete.
But-we
know how to handle noise. In fixed algorithms like speech, it’s that noisy Markov chain stuff. I have to believe that neural network image processing is at least as good and probably better than that.
Admittedly, a one pixel image receiver blinking on and off isn’t going to do much for one. But I strongly suspect that increasing the image resolution is going to fairly quickly reach the point of negative returns, where it takes longer to process the increased number of pixels than any added benefit it might create in determining whether there’s a car up there a quarter mile off.
brake checking is a feature
