Robust Background Segmentation Using Background Models for Surveillance Application

Gaussian Mixture Models (GMM) is a very typical method for background subtraction because it possesses a strong resistibility to repetitive background motion. However when it comes to complex environment, some unexpected situations occur, e.g., when illumination changes, gradually or quickly, segmentation is generated with a poor result. Moreover, this method is not capable of distinguishing shadows of moving objects. In this paper features of intensity and texture information are utilized to eliminate the shadow of moving objects. Integrated with modified Gaussian mixture models by redefining the update criterion, proposed algorithm is adapted to the flexible illumination environment. To validate that the proposed algorithm is robust to apply on surveillance system, we provide a metric with set of variables for evaluation, a comparison had been made between proposal and original GMM, results show the accuracy improvement of models using our updated algorithm. Averagely at least of 34.8% decrease of false alarm rate proves the quality of segmentation has been significantly enhanced and proposal is more competent and stable for outdoor surveillance applications.