Investigating streetscape environmental characteristics associated with road traffic crashes using street view imagery and computer vision

Examining the relationship between streetscape features and road traffic crashes is vital for enhancing roadway safety. Traditional field surveys are often inefficient and lack comprehensive spatial coverage. Leveraging street view images (SVIs) and deep learning techniques provides a cost-effective alternative for extracting streetscape features. However, prior studies often rely solely on semantic segmentation, overlooking distinctions in feature shapes and contours. This study addresses these limitations by combining semantic segmentation and object detection networks to comprehensively measure streetscape features from Baidu SVIs. Semantic segmentation identifies pixel-level proportions of features such as roads, sidewalks, buildings, fences, trees, and grass, while object detection captures discrete elements like vehicles, pedestrians, and traffic lights. Zero-inflated negative binomial regression models are employed to analyze the impact of these features on three crash types: vehicle-vehicle (VCV), vehicle–pedestrian (VCP), and single-vehicle crashes (SVC). Results show that incorporating streetscape features from combined deep learning methods significantly improves crash prediction. Vehicles have a significant impact on VCV and SVC crashes, whereas pedestrians predominantly affect VCP crashes. Road surfaces, sidewalks, and plants are associated with increased crash risks, while buildings and trees correlate with reduced vehicle crash frequencies. This study highlights the advantages of integrating semantic segmentation and object detection for streetscape analysis and underscores the critical role of environmental characteristics in road traffic crashes. The findings provide actionable insights for urban planning and traffic safety strategies.