Extraction of moving human objects from their background using bi-thresholding and boundary evaluation

Abstract

The extraction of moving objects from image sequence of a static scene is an important task in computer vision. However, given an input image and the corresponding background image, due to the intersection of the foreground pixels and background pixels in the measurement space, it is challenging to identify a single threshold that can resolve the ambiguity in a satisfactory manner. In this paper, we propose a novel method for accurately classifying pixels into foreground and background based on two thresholded masks. Firstly, we produce two boundaries by automatically selecting two thresholds. Secondly, points are matched between these two boundaries by using dynamic time warping (DTW) algorithm and boundary segment pairs that are significantly different are identified. Thirdly, shadow, curvature and edge responses associated with each segment pair are evaluated for deducing the final boundary. Experimental results show that substantial improvement can be achieved by using the proposed method. Compared with other change detection methods, the proposed method produces more accurate and pleasing results.