History-Based Road Traffic Anomaly Detection Using Deep Learning and Real-World Data

Abstract

Detecting anomalies in road traffic, such as accidents and traffic jams, can provide various benefits to road users and road infrastructure managers, including optimal route planning, redirecting traffic flows, and reducing congestion caused by traffic accidents. Recently, many history-based traffic prediction deep learning methods have been developed to perform this task. These methods detect anomalous traffic by comparing the current traffic situation with a predicted one based on historical data. This paper investigates the possibility of detecting traffic anomalies using a novel combination of traffic prediction and graph anomaly detection algorithms, both using deep learning, in a real-world dataset of highways near Antwerp, Belgium. It first benchmarks configurations with different time resolutions of prediction algorithms in terms of accuracy. Then, a combined configuration including anomaly detection is benchmarked in terms of traffic anomaly detection accuracy. Furthermore, it examines which traffic features can contribute to anomaly detection e.g. speed, vehicle length. Finally, the entire system is tested on real-world traffic data containing anomalies. The results show a decreased anomaly detection performance when using both vehicle speed and length as features instead of only speed, and an increased performance when using larger time resolutions.