A people counting system based on head-shoulder detection and tracking in surveillance video

Real-time people flow information is very useful for security application as well as people management. This paper presents a counting system which consists of four modules: foreground extraction, head-shoulder component detection, tracking and trajectory analysis. Firstly, in order to reduce computation costs and cope with various complex surveillance situations for foreground extraction, an adaptive components number selection strategy for mixture of Gaussians model is proposed. Secondly, pedestrians are detected by their headshoulders, because this part is less varied and less likely occluded from a downward-slope view. Thirdly, each pedestrian is tracked through consecutive frames using the Kalman filter techniques and cost function. Finally, the resulting trajectories are analyzed to count people entering or leaving the scene. Experiment results indicate that our method can be applied in actual application.