Research On Performance Enhancement Of The Non Rigid Structure From Motion

The technique of non-rigid structure from motion (NRSFM), which use the observed information of image sequence, to establish the spatial projection model and then to estimate the 3D structure, also the relevant motion parameters. As an important direction in the field of computer vision research, NRSFM is widely used in many applications such as face recognition, scene reconstruction and so on. In practical applications, high quality images are sometimes relatively less. In this case, the estimation accuracy of the existing NRSFM is usually significantly decreased. On the other hand, the online learning algorithm which based on the high degree of freedom, may result in low reconstruction accuracy and higher time cost. Aiming at the above problems, we carried out the following two aspects of the research work.(1) A sub-sequence-based integrated algorithm is proposed to deal with the NRSFM problem with small sequence sizes. In the proposed method, sub-sequences are first extracted from the original sequence. In order to obtain diversified estimations, multiple weaker estimators are constructed by applying the extracted sub-sequences to a recent NRSFM algorithm with a rotation-invariant kernel. Considering the fact that the estimations of some weaker estimators may have large errors, the trimmed means of the outputs for all the weaker estimators are computed to determine the final estimated z coordinates. Compared to some existing methods, the proposed algorithm can achieve a higher estimation accuracy, and has better robustness.(2) A new online reconstruction algorithm based on differential evolution algorithm is proposed. First, we choose the front several frames which is next to the new frame to calculate the rigid average shape. Furthermore, we use a number of shape matrixes, which are based on different low ranks to describe the non rigid deformation of the object. At this point, the corresponding reconstruction results can be obtained according to different low ranks. Finally, according to the coarse estimation results, we introduce the differential evolution algorithm for further optimization to obtain the final 3D shape. The experimental results based on a set of widely used experimental data, demonstrate that our proposed algorithm can not only improve the accuracy of the reconstruction, but also greatly save computing time.