Research On Binocular Stereo Matching Algorithm Based On Improved KAZE

Vision is that people are born with a kind of important means of cognition of the world and see the world. Approximately 75% of information people obtained from outside is from the visual system. Just like human, computer access to information is increasingly dependent on its own visual system. With the development of computer vision and image processing technology, two-dimensional image information can not meet the needs of the project any more, and the three-dimensional image can made up for make up for the inadequacy. So the research of computer binocular stereo vision becomes of great meaningful.Based on the study of the basic theory of binocular stereo matching, accroding to the comparison of the two stereo matching algorithms of SIFT and KAZE found that, KAZE has the characteristics of strong robustness under the change of light intensity and higher matching rate under the angle of view, which reflected its advantage in binocular stereo matching. However, KAZE also exists in the running time is too long, resulting in the lack of access to information lag. In order to improve the shortage, thesis optimized the algorithm of feature matching on the basis of the traditional KAZE, which improved the efficiency of stereo matching of binocular stereo vision.In the traditional KAZE stereo matching algorithm, KD tree search strategy is the key to feature matching, so optimizes the feature matching time, which means to optimize the KD tree. Traditional KD tree search efficiency is high in low dimension space, but low in high dimension and feature points detected in KAZE algorithm just based on high dimension space. To solve the problem, a random KD tree search algorithm was proposed. Firstly, according to the reference image and the feature point set of the image to be matched, the KD tree with different directions is randomly generated, which means to rotate the feature points set; Secondly, for reducing the time complexity of feature points set, the rotation of the feature points were set for Householder matrix transformation; Finally, By using hybrid priority search method, the KD tree was generated in parallel.In order to verify whether the improved KAZE algorithm can reflect the effect of stereo matching in binocular stereo vision, different rotation angles, different light intensities and different perspectives images were took to stereo matching tests by using traditional KAZE algorithm and the improved KAZE algorithm. The result shows that: under the premise of detecting the same feature points and image matching rate, the algorithm proposed in this paper shortens the running time and improves the efficiency of the execution. Therefore, the improved KAZE algorithm can not only solve the stereo matching problem, but also improves the efficiency.