| Title |
Predicting High-Risk Areas for Child Pedestrian Accidents Using Machine Learning |
| Authors |
채한희(Chae, Han-Hee) ; 이조은(Lee, Jo-Eun) ; 전은비(Jeon Eun-Bi) ; 이경환(Lee, Kyung-Hwan) |
| DOI |
https://doi.org/10.5659/JAIK.2025.41.8.301 |
| Keywords |
Children; Traffic Accident; Urban Space; Physical Environment; Machine Learning |
| Abstract |
Automobiles play a vital role in urban transportation, but the rapid increase in vehicle numbers has caused several urban problems, including
traffic congestion, accidents, and environmental damage. Among these, the rising number of child-related traffic accidents is a major concern.
Children are especially vulnerable due to limited cognitive development and slower reaction times. Despite ongoing efforts to reduce such
incidents, current policies mostly address responses after accidents occur, particularly in high-risk zones, limiting the effectiveness of proactive
prevention. This study aims to develop a predictive model that identifies high-risk areas for child traffic accidents based on physical
environmental factors. Using 18 indicators related to population, road conditions, land use, and facility accessibility, the model was trained
with data from the Seoul Metropolitan Government. Three machine learning algorithms were tested: Decision Tree, Random Forest, and
XGBoost. Among them, XGBoost delivered the highest performance, achieving an R-squared value of 0.9313. SHAP analysis was used to
interpret the model and identify key contributing factors. The proportion of old buildings ranked as the most influential, followed by
pedestrian crossing density, traffic signal density, and the length of school zones. To assess generalizability, the model was applied to
Busan’s urban and accident data. It showed high predictive accuracy with a score of 0.9327, confirming its effectiveness. This predictive
model offers a practical tool for identifying potential accident-prone areas and supporting proactive child traffic safety strategies, especially in
locations with limited accident data. |