N.T. Nguyen et al. (Eds.): KES-AMSTA 2007, LNAI 4496, pp. 645–654, 2007. © Springer-Verlag Berlin Heidelberg 2007 A Block Based Moving Object Detection Utilizing the Distribution of Noise M. Ali Akber Dewan, M. Julius Hossain, and Oksam Chae ∗ Department of Computer Engineering, Kyung Hee University, 1 Seochun-ri, Kiheung-eup, Yongin-si, Kyunggi-do, South Korea, 446-701 dewankhu@gmail.com, mdjulius@yahoo.com, oschae@khu.ac.kr Abstract. Moving object segmentation in complex scene is the basis for video surveillance, event detection, tracking and development of vision agent in industrial robotics. However, due to presence of camera noise and illumination change, simple background subtraction based techniques are not able to detect moving objects properly. In this paper, we present a novel block based moving object detection method which dynamically quests for both local and global properties of difference image to achieve robustness. Noise model of the difference image is determined analyzing the histogram of difference image and block wise local properties are computed. These local properties are compared with the noise model to extract moving blocks. To remove the stair like artifacts of the segmented result, and to obtain smoothed and accurate boundary, a refinement procedure is employed on the boundary regions of detected moving objects. Experimental results show that the proposed method is robust and achieves better performance in dynamic environment. 1 Introduction Automatic extraction of moving objects plays an important role in computer vision. Diverse applications of computer vision such as video surveillance, video coding, video indexing, distributed artificial intelligence in video analytics and robotics vision in multi agent environment are getting benefits from this research [1][2][3]. Image differencing approach is one of the most common and simplest approaches for moving object detection. It highlights the moving object regions or content changes while suppressing the static background. However, this approach encounters difficulties to discriminate changes occurred due to presence of moving objects from changes occurred due to presence of noise or illumination variations. A popular method to discriminate these changes is image thresholding. However, a low threshold value tends to create false alarm and a high value tends to swamp significant changes between the frames, where determination of a global threshold value automatically is another difficult task [4]. In general, the optimal threshold value is a time varying and content dependent parameter [5]. Therefore, an autonomous surveillance system based on unsupervised absolute value thresholding may be characterized as inadequate [6]. ∗ Corresponding author.