Sub-trajectory anomaly detection based on trajectory segmentation merging and grouping clustering

Trajectory outlier detection can be used to detect possible fraudulent behaviors of taxi drivers while carrying passengers. However, most existing research mainly focuses on the overall trajectory anomaly detection without fully considering anomalous trajectory segments. To address the issue, this paper proposes a sub-trajectory anomaly detection method based on trajectory segmentation merging and grouping clustering. Trajectory vectors are clustered using cosine distance, and each trajectory is segmented based on changes in cluster labels. Second, the trajectory segments are merged according to the speed and mergeable thresholds to get the merged sub-trajectory set. Third, the minimum bounding rectangle of each sub-trajectory is obtained, and sub-trajectories located within it that meet the directional similarity threshold are grouped together. Finally, the algorithm determines whether the number of sub-trajectories and the curve distance of each sub-trajectory are less than specified threshold values. If so, these sub-trajectories are marked as anomalous; otherwise, the density clustering algorithm is utilized to cluster each sub-trajectory group, and the anomalous sub-trajectory whose number of sub-trajectories in the clusters is less than the threshold is identified. Experimental results on real trajectory datasets show that the proposed method outperforms other similar methods in terms of detection accuracy.