Abstract: The widespread use of taxi GPS equipment generates a large amount of trajectory data. The detection and analysis of taxi abnormal trajectory can provide useful support for punishing taxi drivers with fraudulent behavior. For the sparse trajectory of taxis, the anomalous trajectory is detected based on the relative similarity of trajectories. Due to its asymmetry, the traditional density clustering method similar to DBSCAN can not adapt to this situation. Therefore, this paper proposes a density-based RDBSCAN algorithm for taxis abnormal trajectory clustering detection. For the candidate anomaly trajectories obtained by clustering, this paper combines the concepts of trajectory density anomaly value and trajectory length outlier value, and uses evidence theory to synthesize the above two factors to determine the abnormal degree of trajectory, and then obtains the TOP-N anomaly trajectory with the highest degree of abnormality. Using real taxi data of San Francisco, experiments are carried out by extracting the same Origin-Destination (OD) trajectory set. The experimental results show that the proposed method can effectively detect the anomalous trajectory and successfully give the TOP-N anomaly trajectory with the highest degree of abnormality.

Key words: abnormal trajectory detection, taxi trajectory, clustering, evidence theory