Research On Occlusion Boundary Detection Approach Of Depth Image Basedon Machine Learning

Research on occlusion is fundamentally important for the development of most vision technologies. No matter image processing within lower level or logic tasks in upper level, occlusion is involved in most vision research fields such as vision measurement, object identification, scene reconstruction, target tracking, robot grasping as well as automatic assembly. Most of vision technologies would fault out or even fail if the vision system couldn’t take correct treatment measure to deal with occlusion problem. In this case, it has become the most critical technology for vision issues that how to make sure the system can identify occlusion correctly. In this paper, based on depth image, occlusion boundary detection method of visual target is studied by using depth information and machine learning theory.Firstly, the concept, obtaining methods and application of depth image are presented. The concept, classification and boundary of occlusion are introduced as well. This paper also provides a brief introduction to the background information on machine learning to be used.Secondly, two new occlusion related features named eight neighborhood total depth difference feature and maximal area feature are proposed together with their calculation methods, by in depth investigation on the correlation between pixels in depth image and occlusion, combined with existing occlusion related features. On this basis, a new occlusion detection approach based on ensemble learning is proposed which detect occlusion boundary of depth image by strong classifier which generated by many weak classifiers.Thirdly, a new occlusion boundary detection approach without Ground Truth is proposed. First of all, a occlusion related feature with good performance in division named maximal positive vector length feature is proposed.Then the depth image is divided based on threshold segmentation of simple image statistics with proposed maximal positive vector length feature.The result of segmentation is used to train machine learning classifier. At last, occlusion boundary is detected by the classifier.Finally, the feasibility and effectiveness of our method are reported with our experiments. Comparison and corresponding analysis for experimental results are also presented in this paper.