Wednesday, February 9, 2011

Machine vision for pedestrian safety

It’s pretty clear now that cameras and machine vision technology have a big future in automobiles. Back-up cameras, lane departure warning, blind spot cameras and now road sign recognition are all available, but here’s an application I hadn’t thought of: pedestrian detection.

That link takes you to a video on the web site of Car & Driver, where they demonstrate and explain a system that Volvo has under development. This consists of a forward facing camera that watches out for pedestrians about to become road accident victims, combined with a radar system to figure out how far away they are.

I find that an interesting combination of sensors. If you think about it, people come in all shapes and sizes, so I see no obvious way to deduce distance from a single 2D image. The vision system obviously has algorithms that recognize human figures, and I’m guessing it also does some kind of optical flow work to predict whether they are about to become a hazard, then it correlates this with the radar data to figure out if the brakes should be applied.

You machine vision gurus may be wondering why Volvo chose this approach. What about time-of-flight-based systems? Why not two cameras for a stereo view? The problem with the first is that the system has to know it’s a human being in front of the vehicle and not just another car, but that could be achieved with stereo, so is the radar just overkill?

No comments: