Driver Behavior Modelling and Risk Profiling Using Large-Scale Naturalistic Driving Data

Abstract

Driver risk profiling is an emerging scheme in the field of Intelligent Transportation Systems (ITS). Conventionally, a risk score of a driver is calculated on a per-trip basis according to the number of harsh braking, hard cornering, aggressive acceleration, and excessive speeding events. Risk scoring in the academic literature and industry has two main limitations. First, risk scoring has been primitively performed based on a pre-assumption on the risk weights of driving behaviors. Second, the conventional method of risk scoring ignores the individual differences between drivers and the variation in their skillfulness levels.

In this thesis, we tackle the first limitation through the utilization of the Strategic Highway Research Program 2 (SHRP2) large-scale Naturalistic Driving Study (NDS) dataset (i.e., the largest of its kind to date) and performed by Virginia Tech Transportation Institute (VTTI) to develop reliable and robust risk scoring functions. We first utilize the behavioral information of more than 3,000 drivers during crash, near-crash and normal driving events to develop a robust machine learning model that is able to predict the driving risk quantified in terms of crash and near-crash events of drivers given their long-term behavioral patterns. A complete driver profiling framework that considers the joint effect of driving behaviors and environment conditions on driving risk is then proposed and validated. Validation results indicate the robustness of the developed models and framework. Then, a novel safety-based route planner that utilizes the personalized risk profiles of drivers in suggesting individualized routing options is proposed and analysed through a real-world case study that highlights the significance of the proposed route planner.

To tackle the second limitation, we propose a fault inference profiling system in which drivers are profiled based on their individual risk rate. Following the detection of risky events, proposed system can infer the fault contribution of drivers using the time-series radar and vehicular data prior and after risky events. Fault inference is performed through training five customized Hidden Markov Models (HMMs), each representing a behavioral class, on 248 risky events. Promising classification results are obtained and discussed.