A Bayesian Method for Real-time Unsupervised Detection of Anomalous Road Vehicle Trajectories

Anomaly detection is critical in Intelligent Transportation Systems (ITS) due to its significant impact on safety. This paper introduces a Bayesian probabilistic framework for identifying anomalous trajectories without explicitly modeling anomalies reliably. The framework can be adapted according to the sensor quality, balancing speed and accuracy, and avoids out-of-sample performance issues commonly encountered in deep learning methods. By reducing the dimensionality of time series data using Functional Principal Component Analysis (FPCA), a prior distribution of FPCA scores is learned and continuously updated in an online manner. We conducted numerical experiments to validate the method’s effectiveness in detecting common road hazards such as wrong-way driving, over-speeding, and sudden hard-braking. Results demonstrated reliable detection of all tested anomalies with a single detector. Our framework significantly reduced false alarms compared to the Local Outlier Factor (LOF) method, more responsive than Isolation Forest (IF) and successfully mitigated the out-of-sample unpredictability associated with deep learning approaches like VAE-LSTM. Furthermore, it requires low computational resources, making it suitable for implementation across various embedded driving platforms. By addressing the these issues, the method could gain human trust in automated safety systems, accelerating their adoption.