Spatio-temporal Reconstruction For 3D Motion Recovery

With the development of computer technology and hardware equipment,threedimensional coordinate data of human joints has gradually gained the attention of academia and industry and been applied in many fields,such as film and television animation production,sports training,virtual fitting,motion sensing games and gradually penetrated into people's lives.Three-dimensional coordinate data of human joints are usually obtained from color images or depth images by human contour extraction,virtual bone curve extraction,two-dimensional joint location,camera calibration and other algorithms.Due to complex movement or occlusion,similar texture of clothing or soft and wrinkled material and other factors,the obtained data has noise or missing joint coordinates,which will affect the analysis and processing of subsequent applications.Therefore,it is of great value to realize the denoising of three-dimensional human joint motion coordinates.In this paper,a three-dimensional human motion recovery algorithm based on spatio-temporal constraints is proposed,which can effectively reconstruct clean three-dimensional coordinate data of human joints.The results and innovations of the paper are as follows:1.Aiming at the research of human motion characteristics,a spatio-temporal constraint based three-dimensional human motion recovery algorithm is proposed to recover accurate and smooth human motion from the damaged three-dimensional joint motion sequence.Among them,the low-rank term ensures the rationality of human movement,the sparse term ensures the smoothness of human movement,and the isometric constraint term promotes the invariance of bone length in time domain,ensuring the accurate recovery of joint position.Compared with previous three-dimensional reconstruction methods based on two-dimensional image or one-dimensional motion track,this algorithm greatly expands the three-dimensional reconstruction method which can directly restore three-dimensional bone.2.Aiming at the proposed three-dimensional human motion recovery model based on spatio-temporal constraints,this paper proposes the decoupling of non-differentiable terms into simpler subproblems,and uses the Lagrange multiplier method based on iterative method to solve them,and gauss-newton method to solve the nonlinear subproblems.3.Due to the low rank approximation for the operation of a large number of motion frames in the three-dimensional human motion recovery model based on spatiotemporal constraints,all samples need to be accessed to perform singular value decomposition in each iteration of the optimization.Such batch-mode optimization will not only bring a large delay,but also require a large amount of computation and memory resources.In this paper,an online version of the 3d human motion recovery model based on spatio-temporal constraints is proposed.The kernel norm is decomposed by explicit low-rank factorization.The frame-by-frame processing of the input motion sequence is realized by decoupling the kernel norm,which greatly improves the computing and storage efficiency.The experimental results show that the two methods in this paper have better motion recovery effect than the most advanced methods in CMU motion capture dataset,Edinburgh dataset and two Kinect datasets.