Trajectory Anomaly Detection with By-Design Complementary Detectors
Shurui Cao, Leman Akoglu · Society for Industrial and Applied Mathematics eBooks · 2025
Trajectory anomaly detection is critical across a wide range of applications, from traffic control, and wildlife conservation, to public transportation optimization. However, detecting anomalies in trajectory data is challenging due to the diverse nature of anomalies. In this paper, we propose CETrajAD, an ensemble method for trajectory anomaly detection that integrates complementary detectors, each targeting different aspects of trajectory anomalies. Our approach leverages three types of trajectory embeddings—Route, Speed, and Shape—that vary in their sensitivity to length, direction, shape, and speed, enabling the detection of diverse anomaly types. We combine detectors from both the embedding and input spaces and show how their complementary nature improves anomaly detection performance. Through theoretical analysis, we demonstrate the conditions when the proposed ensemble design outperforms traditional ensemble methods. Experiments on multiple real-world datasets, containing both simulated and ground-truth anomalies, show that the proposed model consistently outperforms existing baselines.