Investigating the application of deep learning to identify pedestrian collision-prone zones

Abstract The main objective of this study is to understand the factors that contribute to the frequency of both the total pedestrian-vehicle collisions and collisions that involve pedestrian violations and identify collision-prone areas. The two Full Bayes (FB) macro-level models were applied to historical collision records of the City of Hamilton to identify the collision-prone zones and the key factors that contribute to collision occurrence in TAZs. Finally, a self-organizing map (SOM) deep learning model was developed to identify collision-prone zones for the two collision classes. The results showed that the SOM model identified collision-prone zones with a high accuracy that exceeded the traditional Bayesian approach, based on the developed consistency test. As for the total collisions, the SOM model revealed that intersection density is the most important factor in distinguishing between collision-prone and non-collision-prone zones, followed by the pedestrian network directness and the proportion of residential land uses. As for the collisions that involved pedestrian violations, intersection density was also found to be the most important factor, followed by the density of bike-share stations and parking lots in a TAZ. The results of this study could aid planners in designing pedestrian-friendly networks and develop specific recommendations to enhance safety in unsafe zones.