Evaluation of real time stereo vision system using web cameras

Stereo vision is a computationally expensive process that limits accurate real time performance. However, this is usually due to large resolutions in high end cameras. When low cost web cameras are exchanged for high resolution cameras, a practical real time stereo system can achieve real time performance. The accuracy of the disparity (depth) data depends strongly on the stereo camera calibration and rectification. Chessboards that have landscape orientations, which take at least 50% of the physical image size in a camera and move across the horizontal axis between sample views, give the best data for Levenberg-Marquardt iterative stereo camera calibration algorithm and Bougeut´s stereo rectification algorithm to converge to an accurate result. Our research, using a 7×7 SAD window correspondence algorithm with a modern PC, generates approximately 8 frames per second with a sequential single thread programming model. The frame rate of our system provides reasonably real time response; however, the accuracy of depth data limits the number of applications. Therefore, applications that require detecting relative distances rather than absolute distances between objects, such as wheelchair navigation around buildings, could be equipped with our stereo vision system.