Tracking and segmentation of moving objects in a scene

Problems in computer vision are notorious for being computationally demanding. In many cases a large part of the computation is devoted to interpreting a scene prior to any action being taken. If a response can only take place after the scene has been interpreted, then inevitably that response will be slow and almost certainly not possible in real time. To overcome this problem, it is proposed that a low-level system should be developed, to be included in a general computer vision system. The low-level system would produce 'knee-jerk' responses to external conditions, although these actions could be over-ridden by the higher-level part of the system. This sub-system should be able to do some of the work of responding to the world, and be able to pass information back to the higher-level system, hopefully to ease the computational burden. The solution proposed in the paper is for the sub-system to concentrate on changing pixels only. The assumption is made that changing pixels are intrinsically more interesting than unchanging pixels, and that certain predictions can be made about the scene based on the information derived from the changing pixels.<>