Using Multilevel Hidden Markov Models to Understand Driver Hazard Avoidance during the Takeover Process in Conditionally Automated Vehicles

Wang, Manhua; Parikh, Ravi; Jeon, Myounghoon · 2023 · Crossref

DOI: 10.1177/21695067231192612

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This study addresses the critical safety challenge of ensuring smooth transitions from automated systems to human drivers in conditionally automated vehicles (CAVs). Specifically, it focuses on modeling the continuous hazard avoidance process during the takeover phase, where drivers must regain control to navigate static road hazards. While previous research has largely focused on discrete responses or pre-action reaction times, this work aims to uncover the dynamic, temporal evolution of driving states using computational simulation. The authors propose that understanding these continuous maneuvers is essential for developing robust models of human-automation collaboration and for potentially using hazard avoidance patterns to recognize driving styles. To achieve this, the researchers employed a multilevel Hidden Markov Model (MHMM) to analyze driver behavior data collected from a driving simulator. Twenty-four licensed participants engaged in simulated SAE Level 3 conditional automation scenarios, responding to speech-based takeover requests for static hazards, primarily construction sites. The study utilized manual control input data—specifically gas, brake, and steering inputs—sampled at 10 Hz during the manual control period. The MHMM framework allowed for the estimation of both group-level and subject-specific parameters, accommodating heterogeneity among drivers. The researchers tested three model structures with varying numbers of hidden states: three, four, and five. The hidden states were conceptually defined based on the hazard avoidance process, including phases such as approaching, negotiating, and recovering from the hazard. The results indicated that the three-state model, comprising Approaching, Negotiating, and Recovering, provided the best model fitness, evidenced by the lowest Akaike Information Criterion (AIC) score of 504.16 compared to the four-state (532.70) and five-state (587.76) models. When applied to a testing set comprising 20% of the participants, the trained model achieved an average prediction accuracy of 65.7% in identifying the correct hazard avoidance states. The model demonstrated robustness in predicting the Negotiating and Recovering states across participants. However, the Approaching state proved less stable, often being misclassified as Recovering, particularly for drivers with aggressive styles who maintained higher speeds. The analysis revealed that while the model could effectively capture general patterns, individual driving styles significantly influenced the observability of specific states. The study concludes that hazard avoidance in CAV takeovers can be effectively modeled as a continuous process with three distinct subprocesses. The findings suggest that MHMMs are viable tools for decoding hidden driving states from manual control inputs, offering potential applications in recognizing driving styles and assessing driver readiness. The authors highlight limitations regarding the exclusion of vehicle kinematics and environmental features, such as road curvature, which may affect model generalizability. Future work should incorporate these additional data sources and validate the model across different alert modalities and automation levels. Ultimately, this research contributes to the development of more sophisticated computational models for human-automation interaction, supporting the design of safer takeover processes in automated vehicles.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-nvidia on 2026-08-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• situational awareness
• takeover transitions
• automation
• automation surprise
• mode awareness
• anticipation

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Theoretical Contribution: computational model, conceptual framework, theory or model