An Eigenbackground Subtraction Method Using Recursive Error Compensation Zhifei Xu 1 , Pengfei Shi 1 and Irene Yu-Hua Gu 2 1 Institute of Image Processing and Pattern Recognition Shanghai Jiao Tong University, Shanghai 200240, P.R.China zfxu@sjtu.edu.cn, pfshi@sjtu.edu.cn 2 Department of Signals and Systems Chalmers University of Technology, SE-412 96, Gothenburg, Sweden irenegu@chalmers.se Abstract. Eigenbackground subtraction is a commonly used method for moving object detection. The method uses the difference between an input image and the reconstructed background image for detecting fore- ground objects based on eigenvalue decomposition. In the method, fore- ground regions are represented in the reconstructed image using eigen- background in the sense of least mean squared error minimisation. This results in errors that are spread over the entire reconstructed reference image. This will also result in degradation of quality of reconstructed background leading to inaccurate moving object detection. In order to compensate these regions, an efficient method is proposed by using re- cursive error compensation and an adaptively computed threshold. Ex- periments were conducted on a range of image sequences with variety of complexity. Performance were evaluated both qualitatively and quantita- tively. Comparisons made with two existing methods have shown better approximations of the background images and more accurate detection of foreground objects have been achieved by the proposed method. 1 Introduction Detection of foreground objects in video often requires robust techniques for background modeling. The basic idea of background modeling is to maintain an estimation of the background image which should represent the scene with no foreground objects and must be kept updated frame by frame so as to model both static and dynamic pixels in the background. Moving objects can then be detected by a simple subtraction and threshold procedure. Recently, many background modeling methods for background subtraction have been proposed. Most background models are pixel-based and very little at- tention is given to region-based or frame-based methods. Typical models, among many others, include Kalman filters [1], Gaussians [2], mixture of Gaussians (MoG) [3, 4] and kernel density estimation [5]. Despite being capable of solv- ing many background maintenance issues, pixel-based models are less efficient and effective in handling light switching and time-of-day problems [6]. Rather,