Safe Control Transitions: Machine Vision Based Observable Readiness Index and Data-Driven Takeover Time Prediction
DOI: 10.48550/arxiv.2301.05805
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 critical safety challenge of transitioning control from autonomous vehicles to human drivers during system failures or long-tail events. The authors argue that safe transitions depend on accurately assessing driver readiness before the takeover, predicting the duration of the transition, and evaluating the quality of control afterward. To this end, the study develops machine learning models to predict two key metrics: the Observable Readiness Index (ORI), which quantifies driver alertness, and Takeover Time (TOT). A primary motivation is to demonstrate that these predictive models are robust across multiple in-cabin camera views, expanding beyond the limited angles used in prior research, and to establish objective metrics for post-takeover vehicle stability. The methodology employs a multi-camera framework using four infrared cameras positioned at the dashboard center, dashboard driver-side, steering column, and rearview mirror. The system utilizes a "Top-Down" approach, first detecting the driver and estimating joint locations using Faster-RCNN and HRNet models. These keypoints localize the eyes and hands, which are then cropped and classified using convolutional neural networks (EfficientNet-B3 for gaze zones; ResNet-18 for hand locations and held objects). The extracted features feed into an LSTM model to estimate ORI over a two-second temporal window. Data was collected from two sources: a controlled test track in Iowa with over 100 subjects performing non-driving related tasks, and naturalistic driving data from the LISA-T testbed in California. The authors also introduce two objective ego-vehicle metrics to assess takeover quality: maximum speed deviation ($\Delta v$) and maximum lateral deviation from the lane centerline ($\Delta x$) within five seconds post-takeover request. Results indicate that gaze-based features significantly outperform hand-based features for ORI estimation, likely due to occlusion issues with hands in certain camera views. The dashboard driver-facing and steering wheel cameras provided the most accurate gaze classification, while the dashboard center view was optimal for hand position estimation. The ORI model generalized well across different camera views. Regarding takeover quality, the study found that drivers engaged in highly distracting tasks exhibited higher $\Delta v$ and $\Delta x$, indicating unstable vehicle control. Correlation analysis revealed a slight negative correlation between ORI and both deviation metrics, suggesting that lower readiness predicts poorer takeover quality. Conversely, TOT showed a slight positive correlation with speed deviation, implying that longer takeover times are associated with greater vehicle instability. The significance of this work lies in its comprehensive "before-during-after" analytical framework for control transitions, fully driven by AI. It demonstrates that machine vision can robustly predict driver readiness across varied camera placements, though optimal placement is task-dependent. The findings highlight a tradeoff in camera selection, as no single view optimally captures both eye and hand activity. Furthermore, the introduction of objective ego-vehicle metrics provides a tangible measure of takeover success, linking subjective readiness scores to physical vehicle dynamics. The authors conclude that while correlations are weak, the framework offers a viable path for enhancing safety in conditionally autonomous vehicles by better predicting and mediating control transitions.
Key finding
Gaze-based features provided more accurate predictions of driver readiness than hand-based features in a multi-camera framework, though correlations between these readiness metrics and objective post-takeover vehicle motion remained weak.
Methodology
naturalistic
Sample size: 103
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 author_sweep_intake on 2026-05-28.
| Stage | Outcome | Tool | Model | Prompt | Attempts | Completed |
|---|---|---|---|---|---|---|
| discover | success | author_sweep | — | — | 2 | 2026-05-28 |
| archive | success | canonical_url | — | — | 1 | 2026-06-04 |
| extract | success | cached | — | — | 3 | 2026-06-10 |
| clean | success | clean | — | — | 1 | 2026-06-04 |
| chunk | success | chunk | — | — | 1 | 2026-06-04 |
| embed | success | embed | Qwen/Qwen3-Embedding-8B | — | 1 | 2026-06-04 |
| enrich | success | — | — | — | 1 | 2026-05-28 |
| promote | success | — | — | — | 1 | 2026-06-04 |
| summarize | success | llm | qwen3.6-27b-prismaquant | summ-v5 | 2 | 2026-06-10 |
| tag | success | vector_similarity | — | — | 15 | 2026-06-11 |
| verify | success | — | — | — | 2 | 2026-06-10 |
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.
- takeover transitions
- distraction detection algorithms
- automation
- gaze based attention detection
- situational awareness
- manual
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: measurement protocol
- Theoretical Contribution: conceptual framework