Predicting risky driving behavior with classification algorithms: results from a large-scale field-trial and simulator experiment

Abstract Road safety is a subject of significant concern and substantially affects individuals across the globe. Thus, real-time, and post-trip interventions have gained significant importance in the past few years. This study aimed to analyze different classification techniques and examine their ability to identify dangerous driving behavior based on a dual-approach study. The analysis was based on the investigation of important risk factors such as average speed, harsh acceleration, harsh braking, headway, overtaking, distraction (i.e., mobile phone use), and fatigue. In order to achieve the objective of this study, data were collected through a driving simulator as well as a naturalistic driving study. To that end, four classification algorithms, namely support vector machines, random forest (RFs), AdaBoost, and multilayer perceptron (MLP) neural networks were implemented and compared. In the simulator experiment, RFs and MLPs emerged as the top-performing models with an accuracy of 84% and 82%, respectively, demonstrating its ability to accurately classify driving behavior in a controlled environment. In the naturalistic driving study, RF and AdaBoost maintained robust performance, with high accuracy (i.e., 75% and 76.76% respectively) and balanced precision and recall. The outcomes of this study could provide essential guidance for practitioners and researchers on choosing models for driving behavior classification tasks.