Accumulative difference method for human body motion segmentation in a real-world scene

Abstract

An approach for extracting human body in motion images from a real world scene is presented. It is based on accumulating several difference pictures to recover a moving human body. In difference operations, a technique is proposed to threshold the difference picture, in which an exponential data fitting method is applied to histogram analysis. Then a region-based classification which works on the principle of maximizing the between-class variance is performed to remove the static background that survived. Experimental results show that, owing to utilization of more motion information contained in the sequence, the proposed approach produces a robust and efficient segmentation by only two simple operations of difference and accumulation.<>