Motion Pattern Study And Analysis From Video Monitoring Trajectory

Recently, video surveillance has a wide range of applications in our daily life. As the interdisciplinary of Human Computer Interaction, Computer Vision and Pattern Recognition, video surveillance is currently a highly focused research area of Intelligent Transportation System and visual surveillance technology.This paper presents analysis and research based on trajectory data from video surveillance form structure less crowd behavior scene and complex outdoor scene, which builds a framework including4parts:1. We represent the trajectory group pattern as a Hidden Markov Model (HMM). The hidden states capture the transitive properties of the successive coordinates of the spatiotemporal trajectory, and the sub-trajectories which can be observed as distribution of different states are presented as output symbols. The states sequence that maximizes the probability is the corresponding model for the considering trajectory.2. Trajectory-fusion based on trajectory source, sink point density distribution. For outdoor complex scenes, because of factors such as shadows shade, camera dithering, light intensity change, many tracks were mistaken interrupted. By calculating density distribution between trajectory point and related trajectory, it can be judged that whether the trajectory is abnormal disconnect. For abnormal disconnect point, by comparing the relevant speed direction, space distance and time interval we can implement track fusion to improve the modeling accuracy.3. Pattern learning based on probability fuzzy c-means clustering algorithm (PFCM). Due to the requirements faster models learning speed for outdoor complex scenes, this article introduce PFCM algorithm instead of the traditional fuzzy c-means algorithm. This algorithm not only consider degree of parameters the target relative to the class they belong to, but also the possibility of different groups joined to different centers, thus overcomes the problem of sensitivity to noise and the clustering center of the original algorithm.4. Anomaly Detection based on trajectory model and Behavior Prediction. By comparing the position of the trajectory characteristics and time attributes such as velocity, curvature we can extract abnormal trajectory from input and detect anomaly. Abnormal detection can help us to locate abnormal behavior in video surveillance tracking quickly and monitoring against the dangers of the community. This will be useful in community construction and overtime artificial monitoring, to realize the function of artificial intelligence.