Anomaly Detection on Collective Moving Patterns: A Hidden Markov Model Based Solution

Trajectories of people provide rich information about the collective behaviors, where the term collective behavior means the behavior of a large number of people as a whole. Abnormal people trajectories that are rarely observed may correspond with some unusual events, for instance, natural disasters, terrorism attacks, or traffic accidents. To detect such abnormal people trajectories, namely outliers, this paper presents a solution based on Hidden Markov Model (HMM). The motivation to use HMM for outlier detection on people trajectories is that time-varying people distribution can be modeled by using HMM and HMM can provide the probability that a sequence appears. Here, the time-varying people distributions with low probabilities to appear are regarded as outliers. Experiments with an artificial data set simulating collective behaviors and a real-world traffic data set validate the proposed solution.