The Research Of 3D Non-rigid Body Reconstruction In Trajectory Space Based On Probability Model

In the past decades,NRSFM(non-rigid structure from motion)is one of the research hotspots in machine vision field.However,it is difficult to deal with NRSFM,mainly because different three-dimensional structure shapes will produce similar two-dimensional observation images,while considering the constraint of heavy projection alone is not enough to obtain a single and accurate three-dimensional non-rigid body structure.Therefore,it is necessary to have more priori knowledge about the deformation of real structures and camera shooting motion.This paper adopts the idea of three-dimensional reconstruction based on trajectory space,and combines the probability model to solve relevant parameters.Based on the existing research work,the research contents of this paper are as follows:(1)in this paper,the three influencing factors of the trajectory space-based NRSFM algorithm,namely,the selection of different trajectory basis types,quantities and combination forms,are analyzed qualitatively and quantitatively.Based on the existing auto-selected track basis,the optimal track basis combination and the number of track basis are determined to improve the accuracy of NRSFM.(2)in addition to combining the inherent time smoothness of non-rigid body motion,this paper also considers the spatial smoothness between feature points.The previous formula is difficult to combine with spatial smoothness,resulting in less prior knowledge and poor precision of motion parameters.Spatial smoothness can be represented by a correlation matrix C.In this paper,the correlation matrix C is solved by the precise augmented Lagrangian method(ALM)and the non-precise augmented Lagrangian method(IALM).Different methods are used to solve the correlation matrix C to meet the requirements of the ever-changing non-rigid data sets(3)this paper studies a method to solve the motion parameters of NRSFM based on probability model.It combines the known trajectory space model established by the normal distribution of matrix and converts the information between the feature points in the real three-dimensional structure into prior knowledge.Under the constraint of ADJUST algorithm,the accuracy of prior conditions was improved,and then it was integrated into the probability model in the form of row and column covariance matrix,fully combining the spatial smoothness and temporal smoothness of data set.Under the precise priori conditions,the three-dimensional non-rigid structure with higher precision can be obtained.