Anomaly Detection Based on Trajectory Analysis Using Kernel Density Estimation and Information Bottleneck Techniques

In this paper, we propose a new technique to enhance the trajectory shape analysis by explicitly considering the speed attribute of trajectory data, as an effective and efficient way for anomaly detection. An object motion trajectory is mathematically represented by the Kernel Density Estimation, taking into account both the shape of the trajectory and the speed of the moving object. An unsupervised clustering algorithm, using the Information Bottleneck method, is employed for the trajectory learning to achieve an optimal number of trajectory clusters through maximizing the Mutual Information between the clusters and a feature space of the trajectory data. The trajectories are determined as either abnormal (infrequently observed) or normal by a measure based on Shannon entropy. Extensive tests on simulated and real data show that the proposed technique behaves very well and outperforms the state-of-the-art methods. To the best of our knowledge, this is the first technique to use the Information Bottleneck method on the trajectory data.