A novel framework for detecting non‐recurrent road traffic anomalies by combining a temporal graph convolutional network and hierarchical time memory detector

Abstract

AbstractA non‐recurrent road traffic anomaly refers to a sudden change in the capacity of a road segment, which deviates from the general traffic patterns, and is usually caused by abnormal traffic events such as traffic accidents and unexpected road maintenance. Timely and accurate detection of non‐recurrent road traffic anomalies facilitates immediate handling to reduce the wastage of resources and the risk of secondary accidents. Compared with other types of traffic anomaly detection methods, prediction algorithms are suitable for detecting non‐recurrent anomalies for their potential ability to distinguish non‐recurrent anomalies from recurrent congestion (e.g., rush hours). A typical prediction algorithm detects an anomaly when the difference between the predicted traffic parameter (i.e., speed) and the actual one is greater than a threshold. However, the subjective setting of thresholds in many prediction algorithms greatly affects the detection performance. This study proposes a novel framework for non‐recurrent road traffic anomaly detection (NRRTAD). The temporal graph convolutional network (T‐GCN) model acts as the predictor to learn the general traffic patterns of road segments by capturing both the topological effects and temporal patterns of traffic flows, and to predict the “normal” traffic speeds. The hierarchical time memory detector (HTM‐detector) algorithm acts as the detector to evaluate the differences between the predicted speeds and the actual speeds to detect non‐recurrent anomalies without setting a threshold. In the experiments with traffic datasets of Beijing, NRRTAD outperformed other methods, not only achieving the highest detection rates but also exhibiting higher resilience to noise. The main advantages of NRRTAD are as follows: (1) adopting the T‐GCN with a weighted graph to integrate differentiated connection strengths of multiple types of topological relations between road segments as well as temporal traffic patterns improves the prediction performance; and (2) utilizing a flexible mechanism in the HTM‐detector to adapt to changing stream data not only avoids subjective setting of a threshold, but also improves the accuracy and robustness of anomaly detection.