Matching in dynamic stereo image sequences

Stereo vision algorithms for real-time applications are often characterized by two measures: (i) the speed of the matching algorithm; and (ii) the quality of the output range data as characterized by some error norm. Due to the large number of pixels provided by current imaging sensors, even one iteration of a stereo algorithm over a typical image can be very intensive computationally. Refining the range estimates involves strategies that further increase the computational burden, often requiring hundreds of iterations over a single stereo pair. In this study, we present matching algorithms with polynomial complexity, which provide dense correspondences in static stereo images. They represent a compromise between quality of range data and computational resources. We then extend the methods for dynamic stereo image sequences-allowing for motion of both the stereo platform and objects in the scene. For this extension we use optical flow information from each camera in the stereo configuration to improve the speed and performance of matching. When the calculation of optical flow is impractical, we provide a heuristic alternative which prunes matching graphs without explicit feature tracking. We measure our success through subjective visual assessment of the disparity maps and by calculating the mean-square error between computed disparity maps and the dense ground truth.