Research On Fast Segmentation Based Stereo Matching

Binocular stereo vision fully simulates human eyes to perceive object space distance. It is widely applied in robot navigation, three-dimensional measurement, human-computer interaction and so on. Stereo matching technology is one of the core technology in stereo vision. On the one hand, the performance in textureless and occluded regions still need to improve. On the other hand, stereo matching algorithm needs to adapt to the actual scene and meet real-time requirements. For these two issues, in this thesis, we study the superpixel segmentation, local stereo matching, image segmentation based stereo matching, and ultimately successfully applied these algorithms to detect the road. The main innovations and achievements are as follows:(1) We propose an energy function based real-time compact superpixel segmentation algorithm. A new energy function is defined to make the superpixel with homogeneous color, regular shape and similar size. A coarse-to-fine edge refinement method is proposed to optimize the energy function, avoiding falling into local minima and improving the optimization speed. Finally, we evaluate the proposed algorithm on the Berkeley benchmark dataset. The experimental results show that our method outperforms existing state-of-the-art methods. What’s more, it is able to process a481x321image in real-time on a single core of the Intel i32.5GHz CPU.(2) We propose a new edge-aware dynamic programming based local stereo matching algorithm. Our algorithm performs like optimizing an energy function via dynamic programming. By the edge-aware filter, the pairwise smooth constraint is integrated into the energy function. On the one hand, it makes the disparity discontinuous regions more accurate. On the other hand, it discards the frontal-parallel surfaces assumption used in the traditional local stereo matching algorithm. This allows us aggregate the matching cost at different disparity levels, which makes our method adapt to slanted surfaces. Our algorithm not only outperforms the state-of-the-art local methods on the famous Middlebury benchmark, but also is adaptive to the real-world KITTI vision benchmark.(3) For the segmentation based global stereo matching, we propose a coarse-to-fine disparity optimization method. We combine the new proposed hypothesis testing model and image segmentation information to solve the matching problem in textureless and occluded regions. Traditional segmentation based stereo matching methods restricted the entire superpixel to a disparity plane, so the matching errors caused by segmentation errors can not be corrected. To address this, we segment the image into an irregular pyramid and then relabel the superpixels in a coarse-to-line manner. We correct the segmentation errors by splitting the bigger superpixels and then optimizing the smaller superpixels’disparities. Another disadvantage of the traditional methods is the slow speed. To improve the speed, on the one hand, we apply our fast segmentation algorithm, and on the other hand, we propose an efficient hypothesis testing method for disparity optimization. Experimental results on the Middlebury benchmark show that the proposed method outperforms other state-of-the-art segmentation based methods in terms of accuracy. In addition, the proposed method runs10times faster than the traditional methods.