Expressway Traffic Incident Detection Using a Deep Learning Approach Based on Spatiotemporal Features with Multilevel Fusion

Abstract

The development of an efficient traffic incident detection system is essential for road safety risk warning and active safety control. Despite the fact that a large amount of traffic data can be collected from various detectors installed on expressways, the utilization rate of multisource data is still low, and the spatiotemporal traffic flow data need to be further mined. We propose a multilevel fusion method for the detection of traffic incidents, consisting of both data-level fusion and feature-level fusion. Accordingly, a macro and micro data-level fusion framework was developed, which created virtual detectors by converting video data into virtual loop data, thereby densifying the layout of the original loop detectors without increasing traffic facilities. Both sectional traffic flow data from loop detectors and single vehicle behavior data from video detectors were considered. Based on the fused data, several spatiotemporal variables are constructed to extract the spatiotemporal variation characteristics of traffic flow before and after an incident. The feature-level fusion framework made use of neural networks with a bidirectional encoding strategy to extract multisource features from various detectors and jointly encoded them to generate a comprehensive representation. Specifically, the inner networks extracted features from the multisource data, whereas the outer networks combined features from multiple single-source data to create a comprehensive representation for traffic incident detection. The results demonstrate that the model based on multisource data fusion was superior to single-source models. In addition, the performance of the proposed model was superior to that of the five common methods for detecting traffic incidents. Furthermore, the ablation experiments validated the advantages of key model components.