The Research Of3D Non-rigid Structure From Motion Based On Trajectory Filter

Non-rigid structure from motion (NRSFM) has always been one of the key studies ofcomputer vision and pattern recognition. It refers to the task of recovering the time varying3Dcoordinates of points on a deforming object from their2D locations with image sequences.Shape basis is based on the factorization method, this representation method can successfullyrecovers the3D structure of non-rigid object. But a specific type of shape basis can’t fit allnon-rigid objects situation because of its specificity feature.The duality of spatial and temporal representation lets nor-rigid structure from motionresearch extend to trajectory space, which a moving target can be represented as a set oftrajectory basis combination. According to its own characteristics, trajectory basis can bepredefined, but it can’t simply equivalent that the basis is totally known. Different target requiredifferent basis. So, it is difficult to determine which type of trajectory basis is suitable. The basissize and combination form have great influence on the reconstruction ability. Therefore, thispaper aims at proposing a new algorithm, whether it can completely replace the trajectory basesmethod. In addition, parameters optimization methods directly influence accuracy and efficiencyof the reconstruction algorithm. Whether to find the best solution of rotation matrix becomesanother difficult problem of NRSFM in trajectory space.In view of the above problems, this paper will launch the following works based on existingresearches.(1) In order to solve accuracy and efficiency problem of parameters optimization methods.This paper proposes an improved simulated annealing algorithm (ISA), namely chaoticsimulated annealing algorithm. This method uses chaotic search operator as an operator insimulated annealing algorithm. Based on constrained optimization conditions, by recoveringnon-rigid objects, such as Yoga、Pickup and Stretch. The results show that the operator lowersthe probability of local optimum when SA algorithm in solving complex function optimization,besides, and the method improves the optimization efficiency and effectiveness. By comparing to the LM algorithm, obviously, the solution accuracy has been significantly improved, so thiswork is not only solving the nonlinear optimization problem with outstanding performance, butit’s also a beneficial attempt in terms of improved simulated annealing algorithm.(2) Aiming at the problems from the recovery the structure from motion in trajectory space.This paper proposes a trajectory filter method. The main idea is using alternative methods ofsolving non-rigid structure and motion instead of trajectory basis, arguing that the existingdefinition of reconstruct ability is fundamentally flawed in that it fails to consider systemcondition. The new approach eliminates the need to tune the basis size. Through a large numberof experimental data confirmed: the introduction of trajectory filter methods will be good toeliminate high frequency noise, and realize the reconstruction effect better. Rather than rely onthe experience, or through a large number of data, the work will never try to find which type oftrajectory bases or the basis number etc.