Road Traffic Outlier Detection for Reliable ITS: A Multi-Level Framework from Point to Network Fidelity

Accurate anomaly detection in GPS trajectory data is critical for Intelligent Transportation Systems (ITS), enabling reliable travel time estimation, congestion monitoring, and downstream analytics. Existing approaches, including classical outlier detectors and black-box deep learning models, often suffer from low recall, poor interpretability, or lack of robustness to noisy device streams. This paper proposes a multi-level trajectory cleaning framework that integrates rule-based logic, machine learning, and causal inference for reliable outlier detection in intelligent transportation systems (ITS). The framework operates hierarchically from point-level filtering to trip-level segmentation and network-level map matching and introduces a hybrid Rule-symbolic + causal fusion model that unifies interpretability with predictive accuracy. Unlike previous single-stage detectors, the proposed design ensures cross-level consistency and supports real-time scalability. Experiments on an internal GPS dataset (38 k points) and a public SCATS benchmark confirm high robustness, achieving 97.6% valid-point retention and F1=0.846, AUC=1.0 on public data. The average processing throughput exceeds 800 points/s, confirming suitability for online ITS deployment.