Depth Image Based3D Body Recognition And Occlusion Handling In Real-time Augmented Reality

This thesis proposes a novel algorithm used in depth image based augmentedreality which dealing with human recognition and occlusion. This paper provide anew method to solve recognition problem in augmented reality only based on depthsensor for input data. We develop a series of algorithm such as DVF (Depth VectorFitting), SDL (Static and Dynamic Label) for recognition assistant and MBP(Mapping of Body Parts with Weight) for dealing occlusion. These methods canobviously improve the accuracy and satisfy requirement of real-time comparing toexisted human recognition algorithm in depth based augmented reality solutions.Besides that, our paper also improves existed human recognition algorithm basedon depth image. For the situation with only depth data available, our system is able toimplement the process of rebuilding from raw depth image to final human bodycontour even the list of skeleton nodes. We can obtain the whole accurate backgrounddepth information which supports occlusion dealing and body tracking in subsequentsystem modules.In addition, we focus on the occlusion in human recognition. This papersummaries the situation of occlusion like body self-occlusion and occlusion betweenhuman and other in-door elements. We propose algorithms for both of these two kindsof occlusion based on IK (Inverse Kinematics), Hough, TSDF (Truncated SurfaceDistance Function) and CART (Classification and Regression Tree). Finally, weachieve the purpose of handling self-occlusion and body-environment occlusionwithout any obvious artificial mark in real time.Above all, we set sophisticated experiments for time complexity and recognitionaccuracy comparing with existed algorithms. We proved that our solution has manyadvantages against common depth based recognition methods.