Interactive tracking-based pedestrian segmentation in dynamic scenes

Abstract

Moving object segmentation is highly beneficial to human identification and behavior analysis in intelligent video surveillance. The widely-used background subtraction works not well in dynamic scenes. In this paper, the problem is addressed by first localizing the object by tracking and then segmenting it locally via Graph cuts. We also propose a robust tracker combining the merits of two existing methods [1] and [2], and display an interactive segmentation system. Experiments verify the feasibility of our method and that the proposed tracker outperforms most state-of-the-art methods.