Quantitative Identification of Driver Distraction: A Weakly Supervised Contrastive Learning Approach
DOI: 10.1109/tits.2023.3316203
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 limitations of existing driver distraction detection systems, which typically rely on supervised classification of discrete behavior categories. Such approaches struggle to recognize unseen driving activities and fail to provide continuous distraction metrics necessary for downstream applications like adaptive takeover schemes. To overcome these challenges, the authors propose a vision Transformer-enabled supervised contrastive learning framework that quantifies driver distraction by measuring the distance of specific behaviors from a centralized representation of normal driving in a latent space. The methodology employs a Swin Transformer as the backbone encoder to extract hierarchical feature representations from input images. The model architecture includes a binary classification decoder, a projection head that maps features to a unit hypersphere, and a novel loss function combining binary cross-entropy, clustering-based supervised contrastive loss, and negative log-likelihood (NLL) loss. The NLL loss utilizes a Gaussian Mixture Model (GMM) to cluster normal driving representations, ensuring they are centralized while pushing distracted behaviors away. Distraction levels are quantified by calculating the distance between a given activity’s representation and the normal driving set, using a k-nearest neighbor algorithm for stability. The framework was evaluated on a newly constructed Singapore AutoMan@NTU Distracted Driving (SAM-DD) dataset, collected via a driver-in-the-loop simulator with 42 participants, alongside three other public datasets (AUC-v1, 3MDAD, and DAD). Experimental results demonstrate that the proposed approach achieves superior recognition accuracy and robustness compared to state-of-the-art methods, including various CNN and Vision Transformer backbones. The Swin Transformer encoder consistently outperformed alternatives across all datasets. Crucially, the model exhibited strong capability in identifying previously unseen distracted behaviors, a common failure point for traditional classification models. The quantification of distraction levels was validated through driver skeleton pose analysis, confirming the rationality of the calculated distraction scores. The inclusion of the GMM-based clustering loss significantly enhanced the model’s ability to distinguish unknown distracted activities from normal driving patterns. The significance of this work lies in its shift from discrete classification to continuous quantification of driver distraction, enabling more nuanced inputs for intelligent driving systems. By leveraging weakly supervised contrastive learning and representation clustering, the method reduces dependency on massive labeled datasets of every possible distracted behavior. The introduction of the SAM-DD dataset further contributes to the field by providing high-quality, multi-modal data for evaluating driver monitoring systems. This approach offers a viable, generic technique for real-time driver state assessment, supporting safer human-machine cooperation in autonomous and semi-autonomous vehicles.
Key finding
The proposed supervised contrastive learning framework with Gaussian Mixture Model clustering outperforms existing methods in accurately quantifying driver distraction levels and recognizing unseen distracted behaviors.
Methodology
lab_experiment
Sample size: 42
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 scout_discovery on 2026-05-08.
| Stage | Outcome | Tool | Model | Prompt | Attempts | Completed |
|---|---|---|---|---|---|---|
| discover | partial | scout | — | — | 2 | 2026-05-08 |
| archive | success | unpaywall | — | — | 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 | semantic_scholar | — | — | 2 | 2026-06-04 |
| 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 | partial | — | — | — | 2 | 2026-06-10 |
Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified_with_issues.
Topics
Ranked by relevance to this paper. Hover a topic for its definition.
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).
- Methodological Resource: tool software
- Theoretical Contribution: conceptual framework, computational model