Research And Implementation On Stereo Matching Based On Belief Propagation

Stereo matching is a key technology of computer vision and image processing. It computes a dense disparity or depth map from a pair of images by finding corresponding pixels. Stereo matching is widely applied in3D reconstruction, target recognition, unmanned navigation and so on.In this paper, the background of stereo matching, the principle of stereo vision and the evaluation of stereo matching are first introduced. The works of this paper focus on the Bayes’ theorem and Markov Random Fields. Semi-limited belief propagation is proposed based on belief propagation inference algorithm. The problem is formulated as an energy-minimization framework. The data term is regulated by modifying the cost of unreliable point. Then segmentation cues and occluded cues are incorporated into the prior assumptions to yield the smoothness term. And the belief propagation is break into two phase to apply different prior assumption. All the techniques above help in controlling how message are passed. Then another Hierarchical Belief Propagation based algorithm which successful integrates several recent stereo ideas is introduced. According to the feature of the algorithm, two technologies are explored. Firstly, the matching cost is computed using census transform instead of absolute difference. Secondly, the matching cost is aggregate by cross-base method instead of the time consuming color-weighted method. It will speed up the algorithm but will reduce the accuracy of the algorithm. Experimental results show that the performance of the algorithm in this paper is good.