Robust Data-Driven Framework for Driver Behavior Profiling Using Supervised Machine Learning

Abstract

Driver behavior profiling has been gaining increased attention due to its relevance in many applications. For instance, car insurance telematics and fleet management entities have been recently using smartphones’ embedded sensors, On-Board Diagnostics II (OBDII) units and other on-board IoT devices to collect data on vehicles’ behavior and evaluate the risk profile of drivers. In this context, this paper presents a robust data-driven framework for calculating drivers’ risk profile measured in terms of the additive inverse of the predicted risk probability. The Strategic Highway Research Program 2 (SHRP2) naturalistic driving study (NDS) dataset, which is the largest dataset of its kind to date, is utilized to build the risk prediction models. Crash and near-crash events are used to quantify riskiness whereas balanced baseline driving events (i.e., events captured during normal day to day driving episodes) are used to reflect total exposure or driving time per driver. Thirteen mutually exclusive behavioral risk predictors are identified, and the feature matrix is formulated. A sensitivity analysis is then performed to find the best number of balanced baseline events below which drivers are filtered out. Different machine learning models are selected, customized, and compared to achieve best risk prediction performance. Finally, the utilization of the proposed prediction model within an envisioned driver profiling cloud-based framework is briefly discussed.