Introduction
Fatigue detection is traditionally treated as a binary classification problem—distinguishing between "alert" and "fatigued" states. However, this coarse-grained approach fails to account for the continuous, transitional nature of fatigue. This paper introduces the Multi-Dimensional Fine-Grained (MDFG) modeling framework, which moves beyond binary labels to capture temporal resolution and distributional complexity through synthesized intermediate states and meta-learning.
The Problem with Coarse-Grained Labels
The primary challenge in fatigue detection lies in the limitations of Coarse-Grained Labels (CGL). These labels, often derived from subjective self-assessments like the Karolinska Sleepiness Scale, lack the temporal resolution to capture micro-variations occurring in seconds. Furthermore, they ignore the "intermediate states"—the transition period where a person is neither fully alert nor completely fatigued. Treating these states as two non-overlapping distributions leads to significant misclassification, especially for behavioral subtypes like squinting versus slow blinking.
The MDFG Framework
The MDFG framework addresses these gaps through a three-stage pipeline:
- Preprocessor & Trustworthy Extraction: The system extracts Eye Aspect Ratio (EAR) features and employs a "SmallLoss" mechanism. This identifies "trustworthy" samples by prioritizing those with lower loss during early training, effectively filtering noise from the initial coarse-grained data.
- Fine-Grained Label Generation:
- Categorization (FGC): Clustering and Gaussian Mixture Models (GMM) identify subtypes within alert and fatigued states, establishing "base classes."
- Quantification (FGQ): To represent the fatigue spectrum, the framework synthesizes intermediate states by linearly combining Gaussian distributions from the alert and fatigued base classes. This creates a complete set of labels representing various fatigue levels (e.g., 0.1, 0.3, 0.5).
- Network Learning Stage: A meta-learning module is utilized to handle variability and class imbalance. By learning via pairwise comparisons (determining if one sample is more fatigued than another), the model gains a deeper understanding of underlying distribution shifts.
Experimental Results
The framework was evaluated on two real-world datasets: reRLDD (284 alert, 194 fatigue instances) and reDROZY (183 alert, 322 fatigue instances). Unlike simulated data, these involve subjects experiencing genuine sleep deprivation.
MDFG demonstrated significant performance improvements over traditional binary methods. On average, accuracy increased by 10.0% on reRLDD and 12.1% on reDROZY. Ablation studies confirmed that both the fine-grained labeling and the meta-learning module contributed substantially to these gains.
Furthermore, the model exhibited high robustness. While competing state-of-the-art methods often showed drastic performance drops when switching between datasets, MDFG maintained a consistent F1 score with only a 6.1% variance.
Conclusion
MDFG successfully transitions fatigue detection from a binary "light switch" model to a nuanced "dimmer switch" model. By mathematically synthesizing transitional states and identifying behavioral subtypes, the framework achieves superior accuracy and robustness. Future work will explore integrating multimodal data, such as head posture and physiological signals (EEG/ECG), to further enhance detection capabilities across diverse environments.
Summary updated