Bayesian-Optimized CLAHE for Enhanced Drowsiness Detection in Low-Light Conditions Using Time-Distributed MobileNetV2-GRU Architecture

Driver drowsiness remains a critical factor in road traffic accidents, particularly under low-light conditions where conventional computer vision approaches struggle with poor image quality. This study presents a novel approach combining Bayesian-optimized Contrast Limited Adaptive Histogram Equalization (CLAHE) with a Time-Distributed MobileNetV2-GRU architecture for robust drowsiness detection in challenging lighting conditions. Using the NITYMED dataset containing 128 video sequences, this paper systematically compares three preprocessing strategies: original frames, fixed parameter CLAHE (clip limit=2.0), and Bayesian-optimized CLAHE. The methodology employs Bayesian Optimization to adaptively determine optimal CLAHE parameters based on Perceptual Image Quality Evaluator (PIQE) scores, transforming preprocessing into a task-aware component. Statistical analysis using Wilcoxon Signed-Rank Test demonstrates that the Bayesian-optimized approach significantly outperforms baseline methods, achieving mean accuracy of 93.77% ± 0.0521,F1-score of 93.77% ± 0.0522, and AUC of 97.85% ± 0.0145 across 10-fold cross-validation, with peak performance reaching 98.11% accuracy under optimal configuration (p-values < 0.05 for accuracy and F1-score comparisons). The integration of lightweight MobileNetV2 with GRU enables efficient temporal modeling while maintaining computational efficiency with only 62,449 trainable parameters. Results indicate that adaptive preprocessing significantly improves feature visibility and model convergence, demonstrating practical viability for deployment in Advanced Driver Assistance Systems (ADAS) when implemented with periodic optimization strategies.