Enhanced Driver Profiling Through Clustering of Naturalistic Driving Data for Risk Evaluation and Insights

Driver behavior profiling plays a vital role in understanding patterns that impact road safety and crash risks. The methodology comprises three key phases: exploration, profiling, and validation. In exploration phase, a hierarchical clustering algorithm is used to uncover underlying data structure and assess cluster's optimal number. During profiling phase, clustering algorithms encompassing K- Means, OPTICS, as well as GMM (Gaussian Mixture Models) are applied for precise cluster assignment. The final validation phase assesses the clustering outcomes using performance metrics like the silhouette score. The study analyzed naturalistic driving data from approximately 27,000 trips involving 105 drivers. It identified three distinct driver profiles: less risky drivers, average risky drivers, as well as riskier drivers. Clustering relied on key driving behavior metrics: the frequency of speeding incidents, headway violations, along with harsh events per 100 kilometres. Across all algorithms, the less risky driver profile was consistently identified. K-Means revealed a higher level of aggressiveness among certain drivers, emphasizing dominant driving features. However, the research also highlighted that increasing number of clusters diminishes interpretability of driver profiles. Silhouette analysis indicated that K-Means performed better than OPTICS and GMM as number of clusters increased. Additionally, correlation between driver profiles as well as individual characteristics suggested that aggressiveness is strongly influenced by personality traits. To enhance profiling accuracy, demographic factors encompassing age, gender, as well as history of accidents should be included. The findings of this research have practical implications, offering personalized feedback to drivers to mitigate crash risks and promote safer driving behavior.