Annotated Drowsiness Detection Dataset Captured Using Raspberry Pi 5

Research Hypothesis:
This study hypothesizes that drowsiness can be accurately detected in real-time through computer vision analysis of facial features, specifically eye closure patterns and yawning behavior, using affordable edge computing devices like the Raspberry Pi 5.

Data Collection Methodology:
The dataset was recorded using a Raspberry Pi 5 equipped with a Camera Module 3. All recordings were captured at a consistent frame rate of 30 frames per second (FPS) and a resolution of 640×480 pixels, utilizing H.264 compression for video encoding. The recordings cover various lighting conditions with differing lux levels to simulate real-world scenarios ranging from low-light conditions (e.g., nighttime environments) to bright daylight settings.

Dataset Composition and Labeling:
The dataset includes labeled categories essential for training and testing machine learning models. Labeling was performed using Edge Impulse, with the following categories:

Open Eyes: 968 training samples, 225 testing samples (Total: 1,193)
Closed Eyes: 158 training samples, 49 testing samples (Total: 207)
No Yawning: 496 training samples, 124 testing samples (Total: 620)
Yawning: 60 training samples, 12 testing samples (Total: 72)
Video-wise Annotation Data:
The dataset also contains detailed video-wise annotations for testing purposes:

Open Eyes (mata_terbuka): 30,922 testing samples
Closed Eyes (mata_tertutup): 4,662 testing samples
No Yawning (tidak_menguap): 16,877 testing samples
Yawning (menguap): 1,019 testing samples

Notable Findings:
The data reveals a significant class imbalance, with open eyes representing 85.4% of image samples and 86.9% of video samples, while yawning behavior accounts for only 5.2% of image samples and 5.7% of video samples. This distribution reflects natural human behavior patterns where drowsiness indicators occur less frequently than alert states.

Data Interpretation and Usage:
This dataset can be used to train machine learning models for drowsiness detection applications, particularly in automotive safety systems or workplace monitoring. The class imbalance should be addressed through appropriate sampling techniques or weighted loss functions during model training. The multi-modal nature of the data (both image-based and video-based annotations) allows for both static image classification and temporal sequence analysis approaches. Researchers should consider the lighting condition variations when evaluating model performance across different deployment scenarios.