Multi-Scale Detection of Anomalous Spatio-Temporal Trajectories in Evolving Trajectory Datasets

Chenhao Wang,Lisi Chen,Shuo Shang,Christian S. Jensen,Panos Kalnis

DOI: https://doi.org/10.1145/3637528.3671874

2024-01-01

Abstract:A trajectory is a sequence of timestamped point locations that captures the movement of an object such as a vehicle. Such trajectories encode complex spatial and temporal patterns and provide rich information about object mobility and the underlying infrastructures, typically road networks, within which the movements occur. A trajectory dataset is evolving when new trajectories are included continuously. The ability to detect anomalous trajectories in online fashion in this setting is fundamental and challenging functionality that has many applications, e.g., location-based services. State-of-the-art solutions determine anomalies based on the shapes or routes of trajectories, ignoring potential anomalies caused by different sampling rates or time offsets. We propose a multi-scale model, termed MST-OATD, for anomalous streaming trajectory detection that considers both the spatial and temporal aspects of trajectories. The model's multi-scale capabilities aim to enable extraction of trajectory features at multiple scales. In addition, to improve model evolvability and to contend with changes in trajectory patterns, the model is equipped with a learned ranking model that updates the training set as new trajectories are included. Experiments on real datasets offer evidence that the model can outperform state-of-the-art solutions and is capable of real-time anomaly detection. Further, the learned ranking model achieves promising results when updating the training set with newly arrived trajectories.