Realtime stereo vision processing for a humanoid

This report describes the design of a vision system which is used to control a pair of cameras to move in a way like humans move their eyes. An existing algorithm has been used to give a value of ‘interestingness’ to each region in the image. This algorithm was extended to work in a moving-camera setup. It has been used to control the cameras to attend interesting targets in its environment in a human-looking way. A stereo matching algorithm has been designed to let the two cameras converge such that the attended target is in the center of both camera images, enabling target depth estimation. These two algorithms have been combined resulting in a setup that scans its environment for salient targets and determines their distance.