Unsupervised anomaly detection based method of risk evaluation for road traffic accident

Elevated road plays a very important role as corridors in urban traffic network, and the occurrence of traffic accidents often causes a great impact. In that sense, we propose a unique and enhanced Autoencoder (AE) to identify elevated road traffic accident (RTA) risk based on traffic anomaly detection in an unsupervised manner. An attention mechanism is introduced to extract the traffic condition features considering traffic spatiotemporal variation characteristics. Additionally, an enhanced loss is also introduced to optimize the ability of unsupervised anomaly detection (UAD) approach to detect anomalous RTA risk and persistent anomalous traffic condition, which can significantly boost the anomaly detection performance using the contaminated traffic condition datasets. To assess the RTA risk, the evaluation mechanism and discriminant threshold are used to quantitatively analyze the detected abnormal traffic condition. Finally, experiments on real traffic datasets demonstrate the effectiveness of the model.

Automated summary

In this paper, an enhanced Autoencoder model is proposed to identify elevated road traffic accident (RTA) risk based on traffic anomaly detection in an unsupervised manner. By introducing an attention mechanism and an enhanced loss, the model can effectively extract traffic condition features and optimize anomaly detection performance. Experimental results demonstrate the effectiveness of the model.