Research On The Approach For Occlusion Detecting Based On Depth Image

Occlusion problem has always been important in the field of vision research.Occlusion problem is involved in object recognition,3D reconstruction, target tracking,motion estimation, robot grasping, automatic assembly, spacecraft docking and other areas.It has become the bottleneck problem in the field of visual information processing. Andocclusion problem plays a foundation role for most visual technologies. In this paper,based on the visual target depth image information, the phenomenon of visual targetocclusion detection method is studied by using differential geometry, statistical analysisand machine learning theory.Firstly, depth image definition, depth image obtaining method, occlusion definition,occlusion classification and the relationship between depth image and occlusion aredescribed. And the basic knowledge of machine learning to be used in this paper is alsodescribed.Secondly, maximal depth difference feature and included angle feature are proposedby deep research in the relationship between depth image information and occlusion. Andthe extraction methods are also provided. On this basis, an occlusion detection methodbased on the idea of machine learning is proposed. In this method, Support vector machineis used for detecting the occlusion boundaries in depth image.Thirdly, according to the existing occlusion related features and statistic theory, depthdispersion feature is proposed. And the Gaussian curvature is treated as occlusion relatedfeature at the first time. Then the suitable occlusion features are selected to detectocclusion by analyzing and evaluating all the occlusion related features. On this basis, aRandom Forest based occlusion detection method is proposed.Finally, experimental results show that the proposed occlusion detection methods arefeasible and effective. And the analysis of experimental results is done.