Detection of Discomfort in Autonomous Driving via Stochastic Approximation

Kretzschmar, Florian; Beggiato, Matthias; Pichler, Alois · 2022 · Crossref

DOI: 10.54941/ahfe1002437

archive: archived pipeline: cataloged verified

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

Summary

This paper addresses the challenge of detecting passenger discomfort in automated driving to enhance user experience and enable real-time vehicle adjustments. While physiological indicators like heart rate and pupil diameter are known markers of discomfort, integrating them into a functional multivariate model requires careful consideration of data structure. The study specifically investigates whether classification models should be trained on individual participant data or aggregated across participants, and whether they should account for specific discomfort situations. The researchers utilized data from the “KomfoPilot” driving simulator study, involving 40 participants aged 25 to 84. Each participant underwent two automated driving sessions featuring three near-collision scenarios with a truck. Physiological data—including pupil diameter, interblink interval, heart rate, and head movement—were collected alongside continuous self-reported discomfort levels via a handset control. The authors developed a clustering-classification model based on stochastic approximation. This two-step approach first clustered training data into representatives labeled by majority vote (comfort or discomfort) and then classified test data based on proximity to these clusters. To address class imbalance, discomfort samples were oversampled. Three modeling approaches were evaluated: an individual approach (training on one participant’s data), an aggregated approach (training on all other participants’ data), and a situational approach (training on specific discomfort scenarios from all other participants). Performance was measured using accuracy, precision, and recall. The individual approach failed to generalize, showing high training metrics but poor test performance. The aggregated approach achieved moderate accuracy (67.2%) and recall (64.3%) but suffered from low precision (27.0%), indicating high false-positive rates. The situational approach yielded the highest recall (up to 77.8%) and comparable accuracy, though precision remained low. The study concludes that the situational approach is the most promising method for detecting discomfort, suggesting that models should be trained based on specific traffic situations rather than individual user profiles. This finding implies that situational context is a stronger predictor of discomfort than individual physiological baselines. However, the low precision across all models indicates a need for improvement. The authors suggest future research should incorporate additional features, such as time components, facial expressions, or body movements, and explore larger datasets to refine individual modeling capabilities.

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-prismaquant on 2026-06-09; verification: verified.

Topics

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

• passenger motion sickness comfort
• simulator sickness

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).

• Empirical Findings: physiological data
• Methodological Resource: tool software, validation psychometrics