Fast moving object detection using improved Gaussian mixture models

Gaussian mixture models(GMM) is a widely used approach for background modeling. However, computational barriers have limited its usage in real-time video processing applications. In this paper, we discussed a new update algorithm to achieve the goal of fast detection. Dirichlet prior are introduced to avoid redundant Gaussian components, reducing the computation time of each pixel. Most of the existing GMM based techniques use background/foreground data proportion, which is highly sensitive to the environment, to detect object. To avoid its possible negative effects on segmentation, we use sigmoid function to approximate the probability of Gaussian component belongs to the background and set a threshold for it to segment. Experimental results show this method leads to a faster and a better segmentation than traditional GMM.