Tensor Total Variation Regularized Moving Object Detection for Surveillance Videos

摘要

The classical Background Subtraction (BS) and Moving Object Detection (MOD) problems function on the matrix framework considering each frame as a matrix. This work proposes a method in which the video data is treated as a tensor throughout the implementation and thereby ensuring efficient utilization of the structural properties of the video volume. It also addresses the dynamic background issue (swaying trees, moving water, waves etc.) by solving a tensor optimization algorithm of a convex formulation that is convergent in nature. Moreover, the low-rank property is used to extract the structured part of the scene while Tensor Total Variation (TTV) is incorporated to draw out the foreground part of the emotive surroundings. The excellence of this method lies in the reduced execution time and on the superiority acquired in quantitative evaluation based on F-measure, Recall, and Precision with respect to the state of the art methods.