It is extremely alarming and nearly unbelievable to hear stories how modern cars, approved in all Western markets, use computer vision systems (camera-only) to achieve collision avoidance by applying full braking when image recognition algorithm guesses there is an obstacle present. Having been involved in some robotics years back, my light-professional viewpoint is that cameras cannot be used for automotive collision detection. Not stereoscopic; even less single cameras. Can not. Manufacturers should be legally prevented from pushing such systems into market, because I believe it is fundamentally impossible to design a system which can be proven to improve, not compromise safety.
I would approve systems which use sensor fusion where vision is an assisting factor, but use real distance sensing as well, e.g. ultrasonic or laser time-of-flight / LIDAR.
Randomly applying full braking is something where false positives in rates enough that normal people regularly see this and discuss it in coffee tables is completely unacceptable, by orders of magnitude. But it appears car manufacturers and politicians disagree with me.
The big issue is false sense of security. A well designed system optimized to give as few false positives as possible could increase safety, even if it only detected say 80% of real obstacles, but then if you are going to tell people about this feature, it is going to affect driving patterns, so to compensate, true positive rate must be higher than say 80%, and with cheap-assed and poorly designed sensor technology, that increase comes with the cost of increase in false positives, too.