A novel background model for real-time vehicle detection

In this paper, a real-time background model initiation and maintenance algorithm for video surveillance is proposed. In order to detect foreground objects, firstly, the initial background scene is statically learned using the frequency of the pixel intensity values during training period. The frequency ratios of the intensity values for each pixel at the same position in the frames are calculated; the intensity values with the biggest ratios are incorporated to model the background scene. Secondly, a background maintenance model is also proposed to adapt to the scene changes([1]), such as illumination changes (the sun being blocked by, clouds, or illumination time-varying), extraneous events (a person stops walking, and stay motionless, people getting out of a parked car, etc.). Finally, a three-stage method is performed to detect the foreground objects: thresholding, noise clearing and shadow removal. The experimental results demonstrate robustness and real-time performance of our algorithm.