Identifying Deviations from Normal Driving Behavior

Hananeh Alambeigi, H; McDonald, Anthony D; Shipp, Eva; Manser, Michael · 2022 · ROSA P / Safety through Disruption (Safe-D) University Transportation Center (UTC)

archive: archived pipeline: cataloged verified

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

Summary

This study addresses the safety challenges associated with the transition of control in partially automated vehicles, specifically focusing on predicting driver errors during takeover events. While existing literature explores factors influencing takeover performance, there is limited understanding of how to predict post-takeover driver errors and the optimal data structures, such as time window sizes, for such predictions. The research aims to determine which combination of machine learning algorithms, data sources, and temporal windows best predicts whether a driver will commit an error after an automated driving system failure. The researchers utilized data from a driving simulator experiment involving 64 participants engaged in SAE Level 2 automated driving scenarios. The study focused on unexpected braking events where the automation failed, requiring manual takeover. Data collected included driving performance metrics (e.g., steering, braking, speed), physiological indicators (heart rate, breathing rate, electrodermal activity), and eye-tracking data (glance behavior). Driver errors were defined as failures to complete necessary subtasks or executing them in the wrong order, resulting in 22 error and 100 no-error instances. The team extracted 73 features across various time windows (3 to 300 seconds) prior to takeover. Three supervised machine learning algorithms—decision tree, random forest, and support vector machine with a radial basis kernel—were trained and evaluated using five-fold grouped cross-validation. The results indicated that the random forest algorithm achieved the highest predictive performance, with an area under the receiver operating characteristic curve (AUC) of 0.72 when trained on a 3-second window preceding the takeover. This performance was significantly better than random guessing and outperformed the other algorithms and larger time windows. Feature importance analysis revealed that median speed was the most critical predictor, followed by heart rate, minimum time to collision, and specific glance durations. The study also found that down-sampling data below 10 Hz reduced algorithm performance, suggesting higher sampling rates are necessary for accurate prediction. These findings imply that short-term physiological and behavioral data immediately preceding a takeover are more indicative of driver error than longer-term trends. The identification of specific critical predictors, such as median speed and heart rate, provides actionable insights for developing driver state monitoring algorithms in automated driving systems. By accurately predicting potential errors, these systems could enhance takeover designs, potentially providing targeted assistive technologies or alerts to improve safety during the critical transition from automated to manual control.

Key finding

The random forest algorithm trained on a 3-second pre-takeover window achieved the highest performance (AUC of 0.72) for classifying driver error, with median speed and heart rate identified as the most critical predictors.

Methodology

simulator

Sample size: 64

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. Discovered via bulk_ingest_rosap on 2026-05-23 (6 acquisition events logged).

Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified.

Topics

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

• automation surprise
• telematics crash prediction
• drowsiness detection algorithms
• temporal
• situational awareness
• takeover transitions

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: behavioral performance data
• Methodological Resource: tool software
• Theoretical Contribution: computational model