Background Subtraction using an Active Contour Model

Abstract

Clear background images are useful for various video processing algorithms. Using the fact that non-background objects and their background scenes are not dependent, a contour-based background subtraction algorithm is introduced. The proposed contour-based background subtraction relies on maximizing the joint distribution between the background image and the contour. The proposed algorithm can correctly guide the contour to fit a non-background object even if the initial contour is located far from the boundary. Contour tracking is applied to predict the initial contour of the current frame. Using it results in faster convergent, thus, the initial contour is positioned near the object boundary. An inital initial contour generator is used in the case that the contour tracking cannot find the best estimated initial contour. The algorithm is able to be used without requiring human intervention. The experimental results show that the proposed algorithm has benefits over background subtraction algorithms, the moving average (MA), and the moving average with selective (MAS), with extra computation cost. Comparing selected background images from the video scene, the average of root mean square is 5.65 pixel values, which is much lower than 8.23 from the MA algorithm and 9.72 from the MAS algorithm.

Keywords: Active Contour Model; Background Subtraction; Background-Foreground Model; Probabilistic Approach