Abstract: In order to accurately and rapidly detect vehicles red-light running, this paper presents vehicles red-light running detection algorithm based on gaussian mixture model and GPU. First background image is constructed by mean-time algorithm from the video data during the first red light, and then the background image is used to initialize Gaussian mixture model parameters, and the gaussian mixture model is employed to test vehicle movement. In order to realize the real-time detection, the authors design and implement vehicles red-light running detection algorithm with GPU parallel computing. Experimental results show that it overcomes the problem that initialization parameters in Gaussian mixture model are unreasonably selected in real surveillance video and speeds up the convergence rate of Gaussian mixture model and the improved algorithm proposed in this paper is effective and real time.