Greener view, safer drive: Using repeated field experiments and deep transfer learning technique to investigate impacts of urban road landscapes on driving performance

Abstract

Driving in urban environments is an essential part of urban residents’ daily life. We still know little about impacts of a wide range of greenness on driving performance in real urban environments, after controlling for socioeconomic, demographic, driving record, and other environmental factors. This missing knowledge prevents policymakers and professionals from using appropriate planning and design of green landscapes to create safe driving environments for numerous urban residents. This study aimed to address this significant knowledge gap by using real-world driving experiments. Each of thirty-four residents performed seven driving tasks so 238 driving tasks were completed in total. Each task lasted one hour and followed a randomly assigned sequence. Road greenness and other environmental characteristics were analyzed using deep transfer learning semantic segmentation based on live videos (30 frames per second), recorded by a camera positioned to capture the driver’s eye view. A serial communication technology, known as Controller Area Network bus (CANbus), was employed to continuously measure driving performance using four parameters. A series of hierarchical regression analyses yielded three major findings: First, an increased mean of greenness was associated with improved driving performance, as demonstrated by all four parameters. Second, an increased variation of greenness was also associated with better driving performance in three parameters. Finally, the mean of greenness displayed a stronger positive relationship with driving performance than the variation of greenness in three parameters. The findings imply that both the quantity and quality of green landscapes are critical for promoting driving performance in urban areas.