Enhancing Sustainable Traffic Safety Through Machine Learning: A Risk Assessment and Feature Selection Framework Using NGSIM Data

Precisely assessing driving danger is essential for various applications, including the advancement of autonomous driving systems and traffic engineering decisions. This study presents a driving risk analysis framework based on the Next-Generation Simulation (NGSIM) dataset. First, vehicles were classified into four risk classes using the Fuzzy C-Means algorithm using five key risk indicators. Subsequently, comprehensive driving behavior features representing vehicle movements were extracted and evaluated for both risk class prediction and driving behavior feature selection. A new driving risk score was developed using Spearman’s rho coefficient weights, which reflect the relationship of each risk indicator to risk levels. This score was observed to exhibit an increasing trend consistent with the sequential structure of the Fuzzy C-Means (FCM) clustering based on risk labels, thus confirming that it accurately reflects the labeling process. Furthermore, the findings show that the 26 key driving behavior features selected can predict the driving risk score developed using the XGBoost algorithm with over 85% accuracy. Moreover, feature importance analysis reveals that the following distances and inter-vehicle distance variability are particularly effective in determining driving risk. The study discusses the limitations of driving risk assessment based solely on vehicle dynamics and highlights the importance of developing enriched datasets that include multidimensional data sources such as environmental conditions, infrastructure features, traffic density, and autonomous vehicles in future risk prediction studies. Ultimately, this framework contributes to the development of safer and more efficient transportation systems, supporting environmental sustainability by reducing accident-related congestion and promoting resource-efficient traffic management.