How did the chicken cross the road? By using safe pedestrian infrastructure planned with deep learning methods!

Abstract

Car accidents are a major cause of injury and death in the United States and worldwide. Pedestrians and bicyclists are among some of the most vulnerable road users, meaning that extra precautions need to be taken to protect them. We propose better streetscaping as a solution. We use different types of convolutional neural networks (CNNs) to explore the relationship between streetscaping and safety. Of the CNNs we analyzed, we find that a medium-sized YOLOv8 model is the best performing model for object detection and that UNet is the best performing model for image classification. We find that increased sidewalk and tree coverage is associated with fewer bicyclist or pedestrian crashes. We also find a shared distribution of features between intersections that UNet classifies as “safe” vs. “dangerous.”