GIS-based assessment of pedestrian-vehicle accidents in terms of safety with four different ML models

Abstract

Abstract In this study, both micro and macro level evaluation of pedestrian-vehicle crashes were conducted. Macro-level findings were obtained with GIS-based density analyzes, and critical road segments were determined. The data on road characteristics, traffic characteristics, built environment and land use were collected in 70 critical urban road segments. While conducting micro-level research, commonly used multilayer perceptron and C4.5 decision tree, as well as innovative converted fuzzy-decision model and revised fuzzy-decision model, which significantly reduces the expert judgements on fuzzy models, were used. Significant rules were extracted, and were evaluated from safety perspective. Information gain ratio was used to deal with the black-box structure of machine learning models and to examine independent factors in-depth. The best performance was achieved in revised fuzzy decision model with 68.57% accuracy. The results revealed that land use, parking and peak hour volume have high effect, as well as public transport, speed and road type factors have the greatest effect on pedestrian safety. In the light of the results, various managerial implications such as controlling the density of public transport on main arterials, preventing stop-and-go effects, and monitoring vehicle speeds especially during peak hours were suggested to improve pedestrian safety.