Superpixels based Manifold Structured Sparse RPCA for Moving Object Detection
Résumé
Moving Object Detection (MOD) is a fundamental step in various computer vision and video surveillance systems. Methods based on Robust Principal Component Analysis (RPCA) have often been used for MOD. If the low-rank and sparse matrices are relatively coherent, e.g., if there are similarities between the moving objects and the background regions, and/or when the background is more complicated e.g., dynamic scenes, camera jitter, and lighting conditions, the accuracy of these methods deteriorates. It is because these methods assume that the elements in the sparse component are mutually independent, and thus ignore the spatiotemporal structure of the sparse component. To handle this problem, we propose spatiotemporal structured sparse RPCA algorithm for moving object detection. For this purpose, we incorporate two different manifold regularizations on the sparse component based on the local and global invariance assumption. A spatial and a temporal graph Laplacian regularization is encoded in the form of spectral graph structure. Both graphs are constructed using multiple features extracted from superpixels
computed over the input data matrix. We propose a novel objective function to disentangle moving objects in the presence of complicated backgrounds. We evaluate our algorithm on challenging videos taken from six different datasets, including dynamic backgrounds, lighting condition, and camera jitter sequences. Our experiments have demonstrated excellent results compared to the current methods.