A New Method of Vehicle Activity Perception from Live Video

In this paper, we describe an unsupervised model of activity perception by vehicles trajectories in a visual surveillance scene. We introduce a novel trajectory similarity measure based on for comparing trajectories to cluster them. Then using the result of clustering, a dynamic probabilistic network model is constructed and behavior patterns of normal vehicle´s trajectories are obtained. At last, MAP is used to estimate the parameters of abnormal activity model for abnormal detection. The effectiveness and robustness of our approach are shown by experiments using noisy dataset from real world scene. The experimental results show that the novel method can obtain the categories and samples of normal activity patterns automatically and exactly to establish the normal activity model. The novel method has high reliability and adaptability.