A Framework for Building Comprehensive Driver Profiles

Conventional approaches to modelling driver risk have incorporated measures such as driver gender, age, place of residence, vehicle model, and annual miles driven. However, in the last decade, research has shown that assessing a driver’s crash risk based on these variables does not go far enough—especially as advanced technology changes today’s vehicles, as well as the role and behavior of the driver. There is growing recognition that actual driver usage patterns and driving behavior, when it can be properly captured in modelling risk, offers higher accuracy and more individually tailored projections. However, several challenges make this difficult. These challenges include accessing the right types of data, dealing with high-dimensional data, and identifying the underlying structure of the variance in driving behavior. There is also the challenge of how to identify key variables for detecting and predicting risk, and how to combine them in predictive algorithms. This paper proposes a systematic feature extraction and selection framework for building Comprehensive Driver Profiles that serves as a foundation for driver behavior analysis and building whole driver profiles. Features are extracted from raw data using statistical feature extraction techniques, and a hybrid feature selection algorithm is used to select the best driver profile feature set based on outcomes of interest such as crash risk. It can give rise to individualized detection and prediction of risk, and can also be used to identify types of drivers who exhibit similar patterns of driving and vehicle/technology usage. The developed framework is applied to a naturalistic driving dataset—NEST, derived from the larger SHRP2 naturalistic driving study to illustrate the types of information about driver behavior that can be harnessed—as well as some of the important applications that can be derived from it