Anomalous Human Trajectory Detection Using Clustering Methods

Anomalous human trajectory detection is a critical task in security surveillance in working places. Many studies have been proposed and achieved specific results in abnormal trajectory detection over recent years. In this paper, a framework is proposed for detecting anomalies in human trajectories using clustering methods. In particular, we propose two anomaly detection methods using two different clustering algorithms: spectral clustering-based anomalous trajectory detection (SC-ATD) and DBSCAN-based anomalous trajectory detection (D-ATD). Firstly, clustering methods are used to find normal trajectory clusters in the dataset. Then input trajectory is detected whether it is an anomaly using found clusters. Besides, determining the input parameters of clustering methods is investigated in this work. With spectral clustering, we choose the number of clusters using the WB-index. With DBSCAN, a new cluster quality index (CQI) is proposed to find an appropriate value of the Eps parameter, directly affecting DBSCAN’s quality. The proposed methods are evaluated on a real trajectory dataset: MIT Badge. The results show that both proposed methods detect anomaly trajectories with 77.89 % for spectral clustering and 80.83 % for DBSCAN in terms of F1-score.