Unbiased Anomalous Trajectory Detection With Hierarchical Sequence Modeling

Anomalous trajectory detection plays an important role in the field of trajectory big data mining, providing significant support for identifying drivers traveling at inappropriate speeds and detecting cab fraud. Current studies often use equal-sized grids to represent trajectory points, and they mainly focus on the general shape of trajectories while ignoring the spatial density distribution of trajectories. In addition, existing generative models are biased in learning the patterns of normal trajectories, and the same bias exists in processing labeling information. To address the above two problems, we propose an unbiased anomalous trajectory detection method (HS-UATD) based on hierarchical sequence modeling. Our method constructs a hierarchical structure of the entire spatial region using a quadtree, which captures the location density distribution of the entire spatial region. Our model captures the rich spatio-temporal pattern of trajectories containing spatial hierarchical information and learns the probability distribution of unbiased normal trajectories. We employ both clustering algorithms and anomaly injection techniques to obtain unbiased labeling information, and we define trajectories that deviate from the normal pattern as anomalies. Through extensive experiments on three unbiased, biased and real trajectory datasets, we validate the effectiveness of the method.