Improved post-processing for GMM based adaptive background modeling

In this paper, we propose a new post-processing method for Gaussian mixture model (GMM) based adaptive background modeling which was proposed by Stauffer and Grimson. This is a ubiquitous and successful background modeling method. A drawback of this method is that it assumes independence of pixels and relies solely on the difference between current pixel value and its past values. This causes some errors within the foreground region and results in fragmentation of foreground objects detected. Our method uses relaxed-thresholding and adds foreground edge information in close proximity of detected foreground blobs. The close proximity is obtained as the union of convex hulls of close-by regions which we call the hysteresis region. Our results show that we can achieve increased recall rate with the proposed method without much decreasing the precision of the conventional method.