Online Tree-Structured Trajectory Clustering with Anomaly Detection

Abstract

Trajectory clustering of moving objects has received lots of attention with the demand for surveillance data interpretation. Existing works mostly clustered trajectories in an offline manner, where semantic property, geographic property or both are considered. However, the drawback in these methods is that real time monitoring and surveillance is not possible. In this paper, an online tree-structured trajectory clustering framework spanning from trajectory acquisition to anomaly detection is proposed. The trajectory of moving objects is first retrieved from the preprocessing step of foreground segmentation and object association. Challenging scenarios, where trajectories formation may goes awry, is interpreted and targeted with constraints. With a designated displacement length threshold value, trajectory is partitioned into smaller segments for the clustering purpose. Specifically, the geographic property of each piecewise trajectory is examined and similar trajectories are grouped via Longest Common Subsequence (LCSS) algorithm. Experiments on extensive real world data showed that the approach resulted in a compact tree-structured clustering output with anomaly behaviors representation, suitable for continuous monitoring in the surveillance video.