Image Classification using DenseNet-121 Based on MediaPipe Face Mesh for Real-Time Drowsiness Detection

Abstract

The high rate of traffic accidents caused by driver drowsiness and microsleep highlights the urgent need for reliable driver monitoring systems. However, conventional Eye Aspect Ratio (EAR) methods often fail due to their high sensitivity to changes in head poses and ambient lighting conditions, while standard Convolutional Neural Network (CNN) models impose heavy computational loads on hardware. This study aims to implement and evaluate a real-time drowsiness detection system by integrating the DenseNet-121 architecture with MediaPipe Face Mesh. The proposed method utilizes MediaPipe Face Mesh to isolate the left and right eye Regions of Interest (ROI) independently, using a proportional padding of 35%, which are then classified using a DenseNet-121 transfer learning model fine-tuned in two stages across its last 30 layers. Evaluation was conducted using a custom dataset of 2,000 source images from five subjects, yielding 3,926 eye-region samples after extraction and quality filtering, assessed using a Subject-Independent Leave-One-Subject-Out (LOSO) cross-validation protocol. Across five folds, the model achieved a mean accuracy of 83.22% (standard deviation 13.22 percentage points) and a mean AUC of 0.879 (standard deviation 0.131), with performance variation across subjects found to correlate with inter-subject differences in eye-closure expressiveness, where the two lowest performing subjects also exhibited the lowest AUC values (0.707 and 0.769). The system achieved an average total latency of 198.00 ms per frame, equivalent to 5.1 FPS. These findings indicate that the integration of MediaPipe Face Mesh and DenseNet-121 shows meaningful potential for real-time drowsiness monitoring, while also highlighting the importance of subject-independent evaluation and cross-domain generalization for reliable real-world deployment.