Anomalous Sub-Trajectory Detection With Graph Contrastive Self-Supervised Learning

Some anomalous vehicle trajectories may contain fraudulent behavior or traffic accident information. Existing research mostly starts from a global view, treating the entire trajectory as a detection target to determine whether it is anomalous. However, they failed to fully capture the influence of sub-trajectories on the entire trajectory, thus ignoring some anomalous sub-trajectories. This study proposes an anomalous sub-trajectory detection method (GCSL-ASD) based on graph contrastive self-supervised learning, which identifies anomalous sub-trajectories by mining the supervisory information of trajectory data itself. First, we used map matching to project all historical trajectories onto the road network to achieve more accurate trajectory data representation. Then, to enable the model to transform the detection target from individual trajectory segments to sub-trajectories and thus capture anomalies more accurately, we designed a sub-trajectory aggregation module to aggregate continuous trajectory segments into sub-trajectory. Finally, we used graph convolutional network (GCN) to design the generation module and contrastive learning module to capture sub-trajectory anomalies in attribute space and structure space. We conducted comparative experiments on three real datasets to verify the effectiveness of this method and analyzed the reasons for the driver's local detour behavior.