Research On Local Stereo Matching Algorithms Based On Occlusion Information

Vision is an ancient research subject, and important way for human beings to watch and understand the world. Simulating human visual function through computer is our dreams for decades. As the development of visual physiology, psychology of vision, especially computer technology, digital image processing, computer graphics, artificial intelligence and other subjects, simulation of human binocular vision using computers becomes possible. In recent years, computer vision has become an important research field, made great progress and has been widely used. For example it is useful in automatic tracking and recognition of moving targets, the space robot visual control and autonomous vehicle navigation, etc.Visual computing theory is founded in the early1980s by Marr. The basic theory has made stereo vision become an important branch in computer vision. Binocular stereo vision is a very important form of computer vision. Being in a complex3D scene every moment, we can automatically perceive the distance of all the objects from us through our eyes. The principle of binocular vision is to simulate human binocular imaging system using computers, on the basis of fully understanding the mankind visual theory. More specifically it is to obtain two images of the same object from different positions, and match all corresponding points of the two images. In this way disparity of each point is computed and the three-dimensional geometrical information of the object is obtained.A complete binocular vision algorithm includes three parts:camera calibration, stereo matching, and3D reconstruction. Camera calibration is the process of computing the projection matrix of cameras. Stereo matching is finding all corresponding points in the input images and getting disparity and depth.3D reconstruction is3D modeling of scenes with depth information produced in stereo matching. It can be seen that stereo matching is the absolute core of binocular vision algorithms and it plays a connecting role in the entire process of rebuilding.Stereo matching is the most important and difficult problem in binocular vision, and has become an independent hotspot. In the last two or three decades, stereo matching has been widely concerned and studied by computer vision researchers around the world. Research of this field has formed standard system, and made considerable progress. At present there is a unified platform to assess the results of matching algorithms and give the ranking of submitted algorithms. The platform is a good way to promote academic exchanges of researchers and the development of stereo matching.According to the means of estimating disparities, stereo matching algorithms fall into two categories:local and global. Generally, matching methods include four steps:1matching cost computation,2cost aggregation,3disparity computation or optimization,4disparity refinements. Global approaches rephrase the stereo problem as an energy minimization problem. Local methods find corresponding points by comparing similarities of the windows around them. In general, compared to global methods, local ones get poorer matching results but are simpler and faster, thus local methods are preferred in practical applications especially real-time ones. As stereo vision develops, the boundary between local and global approach is becoming blurred. Recent local state-of-the-art methods based on adaptive support-weight are able to deliver disparity maps comparable to those yielded by global schemes.There are several research priorities and difficulties in stereo matching, such as optical distortion, low textures, repetitive textures and occlusions. Accuracy of matching is directly affected by the occlusion problem. Due to occlusion, not only points within occluded areas are hard to get correct disparities, but also points near occluded regions are affected, which is widely known as Foreground Fattening. Though researchers have done lots of work on handling occlusions, still no good solutions are available. Most methods just simply handle occlusions by some post-processing measures. The measure based on left-right consistency check can’t guarantee good results at any time, and should not be advocated for it is not very elegant. In summary, how to handling occlusion is a key issue and difficult problem that has not yet been solved in local matching methods.This paper focuses on the occlusion problem in stereo matching, and the research goal is to offer new ideas and approaches to handling occluded regions, thus obtaining depth maps with higher accuracy and richer three-dimensional information. This research has important theoretical significance and practical value on the theory of computer vision and practical application such as robot vision, three-dimensional measurement. The main work of this paper includes: 1、According to the core step of local algorithm--cost aggregation, local algorithm of recent years are divided into three categories. This classification clearly shows the evolution of local matching methods, and is able to reflect the future development trend of local algorithms.2、The theory and technical research of occlusion detection; Theoretical analysis and solutions of foreground fattening caused by occlusion; put forward the theory that occluded points should not have supports while computing disparities; then we combine occlusion information with popular adaptive weight, and make a new algorithm based on occlusion-aided support weights.The algorithm provides an approach for seamlessly integrating occlusion handling and the core of local stereo methods. It well overcomes the difficulty that occlusion points occur only in one of two input images, and assigns more reasonable disparity values to occluded points. Experimental results show that the side effects of occlusion are weakened, foreground fattening problem is well solved, and sharp depth edges are obtained. In the entire regions including occluded regions, regions near depth discontinuities, and non-occluded regions, accurate disparity values are achieved.3、Researchers pay more and more attention to disparity refinement because it is able to effectively improve the visual effects and accuracy of the depth map with little extra time consumption. In this paper, the concept of disparity inheritance is proposed, and along with occlusion information they are used to improve traditional refinement measures. The new refinement method is novel, easy to implement, and get sharp depth edges without any post-processing steps.