New Techniques of Foreground Detection, Segmentation and Density Estimation for Crowded Objects Motion Analysis

Abstract

Now video surveillance systems are being widely used, the capability of extracting moving objects and estimating moving object density from video sequences is indispensable for these systems. This paper proposes some new techniques of crowded objects motion analysis (COMA) to deal with crowded objects scenes, which consist of three parts: background removal, foreground segmentation, and crowded objects density estimation. To obtain optimal foregrounds, a combination approach of Lucas-Kanade optical flow and Gaussian background subtraction is proposed. For foreground segmentation, we put forward an optical flow clustering approach, which segments different crowded object flows, and then a block absorption approach to deal with the small blocks produced during clustering. Finally, we extract a set of 15 features from the foreground flows and estimate the density of each foreground flow. We employ self organizing maps to reduce the dimensions of the feature vector and to be a final classifier. Some experimental results prove that the proposed technique is useful and efficient.