Answers are grounded in this paper and cite their sources.
To address insufficient sequential dependency modeling, weak fusion of explicit spatio-temporal behavior features, and limited interpretability in online ride-hailing abnormal trajectory analysis, this article proposes an Abnormal Trajectory Detection and Recognition Model (ATDRM). The main contributions are threefold. First, stacked BiLSTM layers with self-attention represent bidirectional temporal dependencies and differences in the relevance of trajectory segments. Second, learned trajectory embeddings are combined with explicit geometric and temporal indicators for anomaly detection. Third, Ward agglomerative clustering groups detected abnormal trajectories in a behavioral feature space. The resulting groups are interpreted using their feature profiles and representative trajectories. Experiments on five Porto OD pairs show that ATDRM achieves an average Accuracy of 0.984 and an average F1-score of 0.957, which are 1.84% and 4.16% higher than those of the strongest baseline model. The average inference time per trajectory is 0.237 ms, indicating preliminary applicability for online monitoring. The model provides interpretable trajectory information for anomaly screening and operational review in online ride-hailing services.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: