Research On Local Matching Algorithms Of Stereo Vision

Computer vision is an important research topic, which is study for how to make computer get information through visual as same as human, in recent years, the topic has already got lots of research achievements which have been used widely, such as robot obstacle avoidance detection, vehicle navigation, target tracking and recognitionStereo vision is an important branch of computer vision. The research is based on the study of the human binocular vision, according to the principles of human binocular vision, take photos of same target object using cameras which are placed at different locations, and then reconstruct the3D spatial information of these photos through image processing. A stereo vision system includes four primary sections: camera calibration, image rectification, stereo matching and3D reconstruction, and stereo matching are the core content of the system.The target of stereo matching is to get the disparity map of the given reference image. The first task is calculate the disparity of pixels according to find matching pixels from the given stereo pairs, and then convert disparity to gray value on the basis of the ratio relationship, at last the disparity map will be got. The execution of matching algorithm can be divided into four steps:(1) matching cost computation,(2) cost aggregation,(3) disparity calculation/optimization,(4) disparity refinement. Matching algorithms can be divided into two categories including global and local method based on the different steps of algorithm. Global algorithm construct an energy function combined data term with smoothing term, disparity of pixels can be calculated by minimizing the energy function, the steps of global algorithm including step1and3, few method also execute step2; Local matching algorithm construct a matching window around matched pixel, the disparity of pixel can be solved by calculating the similarity between two windows, the steps of these methods including step1.2and3. Step4is a separate process, mainly to optimization the solved disparity map. Generally speaking, the global algorithm has high accuracy, but low efficiency, local algorithm has high efficiency while the lower accuracy than the global algorithm. But with the continuous optimization and improvement of the local algorithm, the accuracy of the local algorithm can achieve the level of global algorithms at current.When take photos from the three-dimensional space, there are some objective factors which affect the photo attributes, such as illumination unbalance, the geometry shape of the taken objects, etc., which have become the problem of stereo matching algorithm. Occlusion and light imbalance is the most important problem for stereo matching research, and researchers of stereo visual have focused on these problems. In this paper, we proposed two algorithms to solve the problem of occlusion and illumination.1. An optimized local stereo matching method based on adaptive weights and occlusion detection. The algorithm proposed by optimizing adaptive weight algorithm, according to assign different weight to occlusion and non-occlusion pixels, reduced the affection of occlusion areas for cost aggregation, thereby reducing the rate of false matching, and also we made effective treatment to disparity discontinuous regions. The proposed method improves the overall performance of matching algorithm. In addition, because the low efficiency is a common problem of local adaptive weighting algorithms, we proposed that taking superpixels instead of pixels as the unit of matching based on the optimized algorithm, which greatly improved the efficiency of our proposed algorithm.2. A local matching algorithm based on mutual information and gradient. According to the research of medical image registration algorithm, we proposed that take mutual information and gradient as the matching principle to calculate cost aggregate, according to the strategy of "greatest optimal", selecting the optimal disparity of pixels. According to the new methods, good disparity map can be obtained even if the matching images have different illumination.Finally, we proposed a combination method by combined the two proposed algorithm. According to the combined method, we can calculate the disparity of images without consider the problem of illumination imbalance, because we can get well disparity map according to the combined method, no matter whether matched images have same illumination.