Characterization for complex trajectory and anomaly detection

In recent years, anomalous event detection has got more research attention and trajectory-based method is becoming popular. However, most researchers view trajectory data as a whole so they lost track´s internal characteristics. Analyzing the trajectory structure will discover much more internal information. In this paper, the improved trajectory structure is proposed and the relative similarity is computed to measure the similarity among trajectories. Then trajectories are clustered based on the trajectory similarity by using spectral clustering and modelled as a series of Gaussians. New trajectories will be matched with the models to detect anomalies. Experimental results proved the validity of the proposed method.