Exploring driving behavior as a latent variable in safety modeling. A preliminary analysis from a driving simulator study.

Abstract

Driver behavior, among other factors, plays an important role in road safety. To better understand it, driving simulator studies and field operational studies and/or naturalistic driving studies are often used for analyzing human factors in road safety. In an attempt to eliminate road fatalities and serious injuries on European roads, the EU is adoping a “Zero Vision” (European Commission, 2011), setting targets and investigating the factors affecting road crashes, including infrastructure, vehicle safety, driver behavior, and emergency response. The on-going EU project i-DREAMS (https://idreamsproject.eu/wp/) aims to develop a tool for monitoring driving safety and bringing in the appropriate interventions (real–time and post–trip gamification). The mathematical modeling of a safety tolerance zone as postulated in i– DREAMS is a complex dynamic problem, benefiting from advanced technologies including OBDII, smart phone application (OSeven), Mobileye (ADAS advanced features using a single camera mounted on the windshield), CardioWheel (physiological based on electrocardiogram (ECG)). The proposed methodology in this paper is a data–driven approach using discrete choice models with safety levels as a dependent variable and driving behavior as a latent explanatory variable in these models; this could be formulated as well as “normal”/“abnormal” driving, or in a broader sense aggressiveness indicator. The developed discrete choice models will be tested under different forms (multinomial, nested, but also ordered); the overall methodology also focuses on how to implement these models dynamically. A preliminary step for this is to cluster driving observations into safety levels, based on several clustering techniques, including k–means, hierarchical clustering, and Gaussian Mixture Models. An application of the proposed methods on a dataset from a previous Greek driving simulator study demonstrated that the optimal number of clusters would be four using k–means (elbow) method. In the hEART 2020 conference, the developed discrete choice models will be presented using the safety clusters showcased in this paper.