Moving Objects Detection and Segmentation In Dynamic Video Backgrounds

Abstract

Moving objects often contain the most important information in surveillance videos. The detection and segmentation of moving objects are the basis for object recognition and intrusion analysis. Gaussian mixture model (GMM) is an effective way to extract moving objects from a video background. However, the conventional mixture Gaussian method suffers from false motion detection in complex backgrounds and slow convergence. A novel approach, which integrates an adaptive Gaussian mixture model with a support vector machine (SVM) classifier, is proposed to detect and segment moving objects in dynamic backgrounds for video surveillance. Each pixel in an image sequence is sorted as a background pixel or a foreground pixel by applying mixture Gaussian method. A block-based SVM classifier is further employed to check each foreground pixel, and it classifies the foreground pixel as a motion pixel or a non-motion pixel. All motion pixels are grouped into moving objects. By utilizing both spatial and temporal information, this integrated method is robust to complex environments. Experimental results show this approach significantly decreases the false motion detection and improves segmentation quality of moving objects.