Stereo matching algorithms often use regularization or relaxation methods to refine estimates of disparity in static images. Unfortunately, the computational requirements of these iterative techniques often preclude their use in real-time systems. Furthermore, most real-time stereo matching systems do not exploit the availability of a disparity map from the previous time step to compute the current disparity. We propose two algorithms for correspondence determination in dynamic stereo image sequences using prior disparity maps to localize, and to speed, the search for matches. The first method uses optical flow estimates in the stereo images to constrain the search space of feasible matchings. The second method uses a heuristic search scheme to prune matching graphs without explicit feature tracking or optical flow computation. Both methods can be applied to existing matching algorithms, to reduce their search space. The resulting improvement in matching is demonstrated in sample images and is quantified through the mean-square error of the computed disparity map compared to dense ground truth. The proposed methods are shown to improve the matching speed (decrease the latency) of five different matching algorithms.
ASJC Scopus subject areas
- Signal Processing
- Computer Vision and Pattern Recognition
- Artificial Intelligence