Towards the Safety-Relevant Dimension of Driver Behaviour: A Dual-State Model

TL;DR

This paper introduces a dual-state model to quantify driver behaviour by estimating the probability of being in a defensive state, enhancing understanding of traffic conflict severity through psychological and probabilistic modeling.

Contribution

It presents a novel latent class Discrete Choice Model capturing driver states as probabilistic mixtures, grounded in psychological theory and validated across diverse driving contexts.

Findings

Defensive state correlates with higher sensitivity to spatial and temporal risk.

The model provides interpretable estimates of driver state probabilities.

Validation confirms the robustness of the neutral state across scenarios.

Abstract

We make a methodological contribution by introducing a new dimension of traffic conflict severity: the probability that a driver is in a defensive state. This behavioural probability reflects an internal response to perceived risk and is estimated using a latent class Discrete Choice Model (DCM) that captures driver behaviour as a probabilistic mixture of two latent driving states: a defensive state, representing heightened caution and collision-avoidance intentions under perceived risk, and a neutral state, reflecting routine driving behaviour under low-threat conditions. The framework is grounded in psychological theory, particularly the triad of affect, behaviour, and cognition. It is also informed by two key concepts. First, that event severity exists on a continuum, rather than being confined to binary categories of safe or unsafe. Second, that drivers perceive risk through a dynamic spatial safety field, one that varies with direction, proximity, and the motion of surrounding road users. Applied to the publicly available rounD dataset, the framework yields interpretable estimates of state membership probabilities. The defensive state consistently reflects stronger sensitivity to spatial and temporal risk, while the neutral state captures context-appropriate yet less reactive driving patterns. Importantly, the paper also proposes a method to assess the quality of the estimated probability of being in a defensive state. Because of the duality between the defensive and neutral states, evaluating the consistency of one offers insights into the reliability of the other. To explore this, a multi-step validation procedure is applied across five data subsets representing different driving contexts, including free-flow and diverging scenarios, to examine how well the neutral state generalises beyond the estimation sample.