Driving Style Clustering using Naturalistic Driving Data

Abstract

Knowledge of driving styles may contribute to traffic safety, riding experience, and support the design of advanced driver-assistance systems or highly automated vehicles. This study explored the possibility of identifying driving styles directly from driving parameters using data from the Strategic Highway Research Program 2 database. Partitioning Around Medoids method was implemented to cluster driving styles based on 14 variables derived from time series records. Principal component analysis was then conducted to understand the underlying structure of the clusters and provide visualization to aid interpretation. Three clusters of driving styles were identified, for which the influential differentiating factors are speed maintained, lateral acceleration maneuver, braking, and longitudinal acceleration. Chi-square test of homogeneity was performed to compare the proportions of trips assigned to the three driving style clusters across levels of each driver attribute (age, gender, driving experience, and annual mileage). The results showed that all four attributes examined had an impact on how the trips were clustered, thus suggesting that the clusters capture individual differences in driving styles to some extent. While our results demonstrate the potential of naturalistic vehicle kinematics in capturing individuals’ driving styles, it was also possible that the identified clusters were classifying mostly drivers’ transient behaviors rather than habitual driving styles. More vehicle parameters and information about road conditions are necessary to obtain deeper insights into driving styles.