Driver profiling: The pathway to deeper personalization

Abstract

Defining driver behavior profiles is an essential step to create unique and tailored driving experience. Depending on the driver information used, two types of profiles can be distinguished. The driver’s preferences in terms of car settings constitute the static profile while his driving behavior constitutes the dynamic profile. In this paper, instead of profiling the driving behavior as one of classical behavior classes (e.g. aggressive, normal, and drowsy), we propose a more personalized approach, where the dynamic profile is represented as a graph, automatically built using the driver’s data. The graph states are learned from the driver data using clustering techniques, namely map matching clustering for GPS data and k-means clustering for kinematics data. To avoid duplicate and overlapping clusters, rules for cluster harmonization are defined. The feasibility of our approach was demonstrated, by applying our framework to two driving datasets, UAH-Driveset and Hcilab. The quality of clustering is later validated using the leave one out cross validation method and two newly defined metrics, namely FDR and RoC, measuring the rate of undiscovered behaviors and the stability of the clusters, respectively. The obtained results show that the proposed approach has a low FDR (an average of 8% for Hcilab dataset and 6% for UAH-driverset).

Keywords

• Driver profile
• Data segmentation
• Driving behavior
• Map matching
• k-means clustering