From GMM to HGMM: An Approach In Moving Object Detection

Background subtraction methods are widely exploited for moving object detection in many applications. A key issue to these methods is how to model and maintain the background correctly and efficiently. This paper describes a foreground detector used in our surveillance system characterized by multiple Gaussian statistics. Compared with the existing methods, our Gaussian mixture model (GMM) differs in model initialization, matching, classification and updating. We propose a fast on-line initialization algorithm to train GMM parameters quickly and correctly. All components of the GMM are classified into three kinds: moving object model, still life model and background model, which is effective for complete detection within a certain period of time. GMMs at different scales are organized in a hierarchical manner to handle sharp illumination changes as well as gradual ones. A convenient way to combine luminance distortion with chrominance distortion is presented for shadow detection in complex scenes. Extensive experimental results are provided to highlight the advantages of our detector.