HySamBS: A Hybrid Sample-based Background Subtraction Method

Abstract

Background subtraction technique is the foundation of video analysis applications. Although many background subtraction methods have been proposed, it is still challenging due to the various nature of video scenes. In this paper, we propose an improved method named HySamBS based on the traditional Self-Balanced Sensitivity Segmenter. This method mainly consists of three framework: 1) modified pixel-wise adaptive feedback, 2) region-wise refinement of segmentation results, 3) frame-wise camera motion compensation. The improved feedback mechanism limits the excessive increasing of segmentation threshold. As the traditional sample-based methods are inefficient for detecting intermittent motion objects, the proposed method extra estimates the initialized reference background image via existent background samples. “Ghosts” suppression can be accelerated with identifying whether motion objects exist in the reference background image. The cancellation of neighborhood diffusion prevents motion objects from being absorbed into background samples. In addition, the effect of camera motion on foreground segmentation is further resisted by feature points matching. Finally, comprehensive evaluation results on the ChangeDetection.net dataset indicate proposed method can adapted to diverse challenging videos scenes, and the overall evaluation metric is competitive with state-of-the-art sample-based methods. The complete source code is publicly available at https://github.com/HuangLian126/HySamBS .