Algorithmic Research On Foreground Segmentation With A Shape Prior

Foreground segmentation is a basic problem in computer vision, and it is also a critical problem in many vision-based systems. It has played vital roles in applications such as human motion capture, video surveillance, human-computer interaction, virtual reality, etc. With the development of the study, many solutions are proposed to extract the foreground from any complex scene accurately, real-timely, the most used methods are inter-frame difference method, background subtraction, optical flow method, active contours and shape-prior based method.For the problem of low contrast grayscale video images segmentation, there is no effective algorithm that can accurately segment foreground object. It is impossible to obtain the precise silhouette of the foreground without other information. So we need to use prior information. But the shape-prior method is just translation invariant, it is cannot handle the local rotation or the changes of part structure of the object silhouette.Based on foreground segmentation with an elastic shape prior, this paper proposed following improvements to enhance the performance.This paper proposes a new energy function based on the turning angle. When estimating the similarity between the prior shape template and the object contour, instead of tangent angles, we use turning angles, which is rotation invariance and translation invariance. The experiment results using the new cost function prove that the new method can handle local rotations.Another improvement of this paper is to handle the case that the part structure of silhouette is changed. Firstly, we extract the object edge from the moving edge and the prior template. Then we can get the distance function which calculates the minimum distance from each pixel to the object edge. Combined with the distance function, the new method using the tuning-angle based energy function can handle the changes of part structure of the silhouette.