Evidence record · UDINUS · SDGs
- Unit
- Rektorat UDINUS
- Period
- 2025
- Goals
- Record
- 301fc6e0-149e-4c55-84c8-f7771534894a
- Published
- 2026-10-30
Universitas Dian Nuswantoro conducted an activity on driver drowsiness detection under low-light conditions, involving Alzami, Farrikh; Naufal, Muhammad; Basuki, Ruri Suko; Winarno, Sri; Al Azies, Harun; Lutfi, Syaheerah Lebai; and Brilianto, Rivaldo Mersis as activity participants. The activity addressed the difficulty faced by conventional computer vision approaches when poor lighting reduces image quality. Its method combined Bayesian-optimized Contrast Limited Adaptive Histogram Equalization, or CLAHE, with a Time-Distributed MobileNetV2-GRU architecture for drowsiness detection in challenging lighting conditions. The activity focused on improving image processing and temporal modeling for the stated detection task.
The activity used the NITYMED dataset containing 128 video sequences. It systematically compared three preprocessing strategies: original frames, fixed-parameter CLAHE with a clip limit of 2.0, and Bayesian-optimized CLAHE. Bayesian Optimization adaptively determined CLAHE parameters using Perceptual Image Quality Evaluator scores. This transformed preprocessing into a task-aware component. The detection architecture combined lightweight MobileNetV2 with GRU to enable temporal modeling while maintaining computational efficiency. The reported architecture had 62,449 trainable parameters.
Statistical analysis used the Wilcoxon Signed-Rank Test. Across 10-fold cross-validation, the Bayesian-optimized approach achieved mean accuracy of 93.77% ± 0.0521, an F1-score of 93.77% ± 0.0522, and an AUC of 97.85% ± 0.0145. Peak performance reached 98.11% accuracy under the optimal configuration. Accuracy and F1-score comparisons reported p-values below 0.05. The results indicated that adaptive preprocessing improved feature visibility and model convergence. The activity also reported practical viability for deployment in Advanced Driver Assistance Systems when periodic optimization strategies are implemented. The source presents this UDINUS activity as research on a driver drowsiness detection method for low-light conditions.