Towards the spatial analysis of motorway safety in the connected environment by using explainable deep learning

Abstract

This study investigates the application of Connected and Automated Vehicles (CAVs) as moving sensors that transmit their speed and position in real-time for spatial analysis of motorway safety. Those data are used for the generation of image-alike inputs which describe the speed distribution over the entire motorway model in the form of heat-maps. Their labels are safety categories computed by using average Time-to-Collision (TTC). The Convolution Neural Network (CNN) is proposed to predict the category of safety based on the image-alike labeled dataset. Furthermore, Explainable Artificial Intelligence (xAI) is used to explain which segments of image-alike inputs are critical for the accurate prediction of safety. It is applied to selected inputs with the best learning performance and if they represent the undesirable safety categories. The study investigates the impact of various penetration rates of CAVs with the Intelligent Speed Adaptation (ISA) system on the spatial distribution of safety–critical regions. The higher penetration rates of the CAVs with the ISA system reduce the dispersion and intensity of critical regions computed by xAI over the entire motorway. Those regions are located at the most critical part of the analyzed motorway segment where the on-ramps flow interacts with the mainstream flow and its adjacent off-ramp. The higher penetration rate of CAVs with the ISA system induces a more consistent and localized distribution of critical regions regarding safety. Thus, this confirms that critical regions for safety categorization computed by xAI correspond with the motorway region with the most critical safety.