Spatiotemporal multi-feature fusion vehicle trajectory anomaly detection for intelligent transportation: An improved method combining autoencoders and dynamic Bayesian networks

With the continuous development of intelligent transportation systems, traffic safety has become a major societal concern, and vehicle trajectory anomaly detection technology has emerged as a crucial method to ensure safety. However, current technologies face significant challenges in handling spatiotemporal data and multi-feature fusion, including difficulties in big data processing, and have room for improvement in these areas. To address these issues, this paper proposes a novel method that combines autoencoders, Mahalanobis distance, and dynamic Bayesian networks for anomaly detection. Autoencoders, as powerful unsupervised learning tools, are used for feature extraction and fusion, allowing for a more comprehensive understanding of vehicle behavior, which is essential for identifying anomalies. The Mahalanobis distance-improved dynamic Bayesian network further enhances the model’s detection accuracy and robustness for time series data, improving the efficiency of large-scale data processing and significantly enhancing the ability to fuse and analyze spatiotemporal information. The primary motivation of this research is to improve the detection capabilities of intelligent transportation systems for vehicle trajectory anomalies, thereby strengthening traffic safety. Experimental verification shows that the proposed combined model performs excellently, with significant improvements in detection accuracy. This research not only enhances existing anomaly detection technologies but also provides strong technical support for future intelligent transportation systems, ultimately contributing to overall road safety and reducing traffic accident rates. Additionally, the practical implications include reducing traffic congestion and environmental impacts, making urban transportation systems more efficient and sustainable.