Shape prior based moving object detection

摘要

In intelligent video surveillance, moving object detection plays an important role as the basis of activity understanding and gait recognition. Existing background subtraction post-processing methods don't use shape prior information and previous background subtraction results to guide current detection. As a result, one moving object splits into many connected components. Background pixels are included in foreground object in dynamic scenes. The proposed moving object detection method embeds the shape prior in a Markov random fields (MRF) energy function to improve the result of background subtraction. Gaussian mixture model (GMM) background subtraction is used to obtain the initial results, then MRF energy function with shape prior is used to correct the results. Iterated Conditional Modes (ICM) method is employed to optimize the energy function with the output of GMM as the initial labels. The demonstration shows that the proposed method performs better than GMM and GMM with MRF in dynamic and some other complex environments.