Object Detection By Recovering Sparse And Low-rank Components Of Matrics

Background modeling is crucial to the moving object detection; however, thetraditional background models have complex calculation and lower segmentationaccuracy. It is difficult to cope with complex scenarios such as illumination variation,dynamic textures or other factors. In addition, background modeling becomesnumerically challenging for high dimensionality data. Recently, learning with sparsityhas caused more attentions in machine learning and computer vision research, e.g.principal component analysis (PCA).John Wright et al. propose Robust-PCA (RPCA) to solve these existing problems,and RPCA is a substantial improvement over the classical PCA which can onlycomplete exact recovery from mildly corrupted observation data. Then the researchershave proposed recovering methods such as augmented lagrangian method (ALM), PCP,et al. However, this recovering problem is only seen as a semi-definite programming(SDP) problem to deal with but neglects the separable structure in both the objectivefunction and the constraint.Based on these, several principle similarities between each frame are kept toform a subspace that models global variations. The new frame except for foregroundparts is projected into the subspace, and then the moving objects are left. Thus, thisidea has a huge advantage to handle the global variations in the background such asillumination changes and dynamic textures，and the main work of this thesis is that:The thesis analyzes Robust-PCA, and low-rank and sparse matrix theory isapplied to the problem of moving object detection. And it shows that aboveRobust-PCA problem can be addressed in the new framework, which the observedvideo matrix is decomposed into the low-rank matrix and the sparse matrix. Theoptimization problem can be solved by the proposed method according to matrixrecovery theory. And the experiments prove the proposed framework and method havea good performance.The drawbacks of existing object detection methods are that performance can beeasily affected by complex scenarios such as illumination variation, dynamic texturesor other factors. The optimization problem can be solved by the proposed variant ofalternating direction method (VADM), which accomplishes recovery by taking full exploitation of the separable structure property of the model. The algorithmconvergence of the proposed VADM is derived. Moreover, VADM not only avoids thedefects of convergence of the state-of-the-art algorithms, but also improves thedetection rate of object and reduces the rate of false positives. The experiments verifythat proposed algorithm outperforms the state-of-the-art approaches, and has theattractive robustness and high accuracy of the VADM for illumination variation anddynamic texture.