Research On The Performance About Stability Of Non-rigid Structure From Motion

In recent years,3D reconstruction has become an important branch in field of computer vision, and it is widely applied to many areas, such as medical image processing, diagnosis of mechanical products, robot navigation, etc. In such 3D reconstruction algorithms, non-rigid structure from motion has been widely researched and applied because it’s convenient and cost-effective. Non-rigid structure from motion recovers the 3D structure of the deformation object and information of its movement from 2D image feature point sequences.There are some algorithms of non-rigid structure from motion model currently perform good, but they also have some problems. On the one hand, part of frames in 2D image sequences we get has incomplete feature point information, due to various factors such as cover, rotation when object moves. In that case, we should estimate the missing values first in depth value estimation. The existing missing value estimation algorithm is not ideal to estimate these missing values in the condition of small samples. On the other hand, the result of 3D structure from motion will be changed accordingly when parameters of algorithm change, this causes the algorithm without stability and 3D structure from motion unsatisfactory.This paper aim at these problems carries out ways in two aspects as follows:One is that we propose an integrated model of missing value estimation based on subspace sample according to column space fitting method of non-rigid structure from motion model, to improve the accuracy of missing value estimation. We extract subsequences from the original image sequences firstly. We then use CSF algorithm to estimate the missing data of the subsequences. Finally, the final estimate value can be obtained using linear weighted method to integrating these estimated results. The other is that we propose an integrated method based on rotation invariant kernels algorithm of non-rigid structure from motion model, to improve the stability of the original rotation invariant kernels algorithm. We first set the parameters which affect the reconstruction result. Then, we use NRSFN-RIKs algorithm to reconstruction in each pair of parameter values. When obtain the estimated 3D coordinate values of these points, we can get the final reconstruction results by using average truncated method to integrate these estimated results.Both integration algorithms in this paper are tested on multiple sets of widely used image sequences, the experiment results show that our proposed algorithm is effective to improve the current problems in these existing method, and also improve the reconstruction effect of non-rigid structure from motion model.