Large-scale sensing for driving regimes using in-vehicle data: unsupervised and supervised approaches

Abstract

Advances in sensing technologies have enabled the collection of high-resolution, naturalistic data. Past studies have used data-driven approaches to identify driving patterns, but often overlook how contextual factors shape the operating conditions that define individual driving behaviour. This paper proposes a generalised framework for identifying and reproducing context-dependent driving regimes using integrated vehicle, trajectory, road-network, and meteorological data. We characterise driving regimes based on ten features reflecting behavioural (acceleration and jerking violations, speeding and energy/fuel use), driving state (acceleration, deceleration, and idling), and contextual (urban roads, rain, and visibility) factors. We apply hierarchical clustering and k-means to identify driving patterns across diverse scenarios, then use supervised learning (ANN, SVM, and RF) with a driver-independent hold-out validation strategy to reproduce the cluster-derived regime labels. The database comprises 3.11 million seconds of engine, kinematic, and meteorological data from 2,342 trips by 35 drivers, collected between 2023 and 2026. We observe that driving regimes are shaped by the interaction between behavioural and contextual factors, rather than individual driver characteristics alone. We identify four distinct driving regimes: urban low visibility, urban-rain-low visibility, rural-highway, and urban-normal. All supervised models achieve balanced accuracies and F1-scores of approximately 0.98, thereby demonstrating strong agreement with the data-driven cluster labels. Shapley Additive exPlanations further shows that contextual features dominate model decisions across regimes in this large-scale naturalistic dataset. These findings support context-aware applications such as risk-sensitive insurance pricing, smarter advanced driver assistance systems, and energy-range optimisation.