Improving Gaussian Mixture Model based Adaptive Background Modeling using Hysteresis Thresholding

Background modeling based object detection has a significance in real time surveillance video applications. The method proposed by Stauffer-Grimson is a widely accepted successful method. But in this method, some single piece foreground objects are detected as many separate object pieces. In this study, a hysteresis thresholding method using the union of convex hulls of closely positioned binary connected components in foreground is proposed. In addition, information on temporal edge changes for the foreground is integrated to the model. Consequently, using hysteresis thresholding prevents falsely detecting single foreground objects as many separated smaller objects, and using the information on edge changes for the foreground enhances the performance of foreground detection.