Dynamic Images

A video-recording of a football match shows figures moving around in ways which are, for short periods, relatively predictable. Using temporal information to identify objects in a time series of images is possible using such techniques as differencing consecutive images to distinguish the moving objects from the background. It is, of course, very simple to difference consecutive images and simply measure the amount of change. A simple intruder alarm which accomplishes this is currently obtainable for under fifty dollars. Unfortunately it counts newspapers blown by the wind and passing cars as intruders. Actually being able to distinguish between a man, a sheet of newspaper, and a passing car is possible by the techniques described. More sophisticated recognition issues, such as distinguishing a man from a dog, naturally arise. Because of the huge variety of images which a human being can confidently label as `man' or `dog', the existing methods are not satisfactory. It is possible to submit the whole image to a huge neural net, and this has been tried by optimists in need of an education, but any passing statistician would give a pitying smile if informed of this plan. It is something like trying to recognise text a page at a time because it is hard to segment into letters. You'd have to be pretty desperate to try it.