Abstract:

In view of the low accuracy problem in binocular stereo matching caused by image texture lacking, occlusion and depth discontinuous, this study proposes the matching cost calculation of local image features based on deep learning Siamese neural network model which takes mean filter preprocessing and SLIC super-pixel segmentation as input. The method effectively improves the accuracy of edge regional identification when disparity computation happens, and avoids the problems of edge expansion, blur and discontinuity, and thus improving the accuracy of stereo matching. The experiments are based on the Middlebury stereovision data set test platform, and compared with the disparity maps obtained by SGM, adaptive weight AD-Census and other possible methods. The results show that the algorithm has a significant improvement in the matching effect of the depth discontinuous area and the lack of texture area, and the average parallax error and average error rate are reduced, indicating that the proposed method has better robustness.

Key words: binocular vision, super pixel segmentation, SLIC algorithm, stereo matching, siamese neural network, SGM algorithm, AD-Census algorithm, Middlebury dataset