Research On Performance Enhancement Of Non-Rigid Structure-From-Motion Algorithm For Small-Size Image Sequences

As one important approach of 3D object reconstruction,non-rigid structure from motion(NRSFM)can simultaneously estimate 3D object reconstruction and the corresponding motion parameters,by utilizing the feature points of 2D image sequence.When there are only a few samples in the image sequence,i.e.small-size image sequence,the 3D reconstruction accuracy generally decreases significantly for the existing NRSFM algorithms.The reason is that,a small amount of samples can only establish a limited number of equations.It is difficult to derive the optimal solution when the number of unknown variables is relative large in the equation set.Moreover,it is an intractable problem to extract accurate shape bases via the dictionary training model,when NRSFM is used for single-frame image reconstruction.To address the above issues,two works have been carried out in this thesis.(1)A 3D estimation method is proposed for small-size image sequence based on integrated trajectory group.First,a certain number of overlapping data groups are derived by extracting a set of trajectory groups using the distance weights of the paired points.A large number of small random trajectory groups can increase the viewing angle and reduce errors caused by complex deformation.Then,in order to improve the estimation accuracy,an adaptive rank selection strategy is designed to select the approximate optimal rank parameters.For each trajectory group,the z-coordinates of the observation matrix are estimated by a column space fitting(CSF)algorithm.For the output of the CSF based weak estimators,the final three-dimensional shape is obtained via an alternating directional method of multipliers(ADMM).The experimental results on several widely used image sequences demonstrate the effectiveness and feasibility of the proposed algorithm.(2)A single-frame image reconstruction algorithm is proposed based on an improved dictionary learning method.In the proposed method,a weighted sparse model is designed to obtain more accurate shape bases.In order to further improve the performance of 3D shape estimation,a two-stage scheme,i.e.a main 3D shapeestimation and a compensatory 3D shape estimation,is proposed for 3D shape estimation of non-rigid object.The weighted sparse representation model is adopted to extract the shape basis of each estimation stage.The main 3D shape is estimated by the original observation.The compensating 3D shape is calculated using the residual error between the original observation and the estimated value.Finally,the optimized objective function is constructed via an augmented Lagrangian.The final 3D shape is derived by optimizing the unknown parameters via the proximal operator of the spectral norm.The experimental results on the widely used CMU image sequence demonstrate the effectiveness and feasibility of the proposed algorithm.