A Robust Environment-Aware Driver Profiling Framework Using Ensemble Supervised Learning

Driver profiling is the real-time process of detecting driving behaviors and computing a driver’s expected risk based on detected behaviors. Predicting risk based solely on the inclusion of detected behaviors may not be accurate because this method of predicting ignores the environmental (e.g., weather conditions, traffic density level) context of detected behaviors. Moreover, coupling detected behaviors with their environmental context can be leveraged towards creating personalized risk profiles for drivers in each driving environment. These profiles can be utilized in various ITS applications including personalized safety-based route planning. In this paper, a novel driver profiling environment-aware framework is presented. In the proposed framework, data processing is distributed over three computational layers to enhance the overall reliability of the system. A risk prediction model is hosted on the edge/fog to determine the driving risk while considering the joint effect of the in-vehicle detected behaviors and their environmental context. Risk values along with a driver’s compliance to warnings are both utilized to compute the risk profile on the cloud. Using SHRP2 Naturalistic Driving (ND) dataset, the development of a novel risk prediction model is presented herein with the underlying sub-processes of data preprocessing, error analysis, and model selection. Then we analyze both the performance of the developed risk prediction model and the overall performance of the proposed system. Validation results for the developed model indicate a good compromise between bias and variance. Moreover, the results of the overall risk scoring model reflect its robustness and reliability in assigning accurate risk scores.