Annotated drowsiness detection dataset captured using Raspberry Pi 5

TL;DR

This paper introduces a new dataset for detecting drowsiness using affordable hardware, capturing images under different lighting conditions to improve real-world safety applications.

Contribution

The dataset uniquely combines edge computing optimization with varied lighting conditions, addressing a gap in real-world drowsiness detection.

Findings

The dataset includes 33,750 annotated images across five lighting conditions captured using a Raspberry Pi 5.

Baseline performance using Edge Impulse’s FOMO algorithm achieved 92.8% accuracy for eye state detection and 89.5% for yawning detection.

The dataset reflects realistic drowsiness patterns with a natural class imbalance and includes metadata for training lightweight models.

Abstract

Drowsiness-related accidents represent a critical safety concern in transportation and workplace environments, necessitating real-time monitoring solutions deployable on affordable hardware. This paper introduces the Annotated Drowsiness Detection Dataset, which uniquely combines edge computing optimization with varied lighting conditions (0–615 lux), addressing a critical gap in real-world deployment scenarios. Our dataset comprises 33,750 annotated images collected from 32 participants across five distinct lighting conditions, capturing various states of alertness and drowsiness. Captured using a Raspberry Pi 5 equipped with Camera Module 3, the dataset encompasses facial feature analysis focusing on eye closure patterns and yawning behavior. The recordings were captured at 30 FPS with 640 × 480 resolution using H.264 compression across five lighting conditions (0, 28, 45, 86, and 615 lux) representing nighttime to daylight scenarios. The dataset was systematically collected using standardized protocols, ensuring compatibility with edge computing constraints while maintaining sufficient quality for computer vision applications. The dataset comprises 2092 labeled images divided into training and testing sets—Open Eyes (968/225), Closed Eyes (158/49), No Yawning (496/124), and Yawning (60/12)—alongside 53,480 labeled frames extracted from video recordings: mata_terbuka (30,922), mata_tertutup (4662), tidak_menguap (16,877), and menguap (1019). We provide baseline performance analysis using Edge Impulse’s FOMO (Faster Objects, More Objects) algorithm, achieving 92.8 % accuracy for eye state detection and 89.5 % for yawning detection under optimal conditions, while maintaining 76.8 % accuracy under challenging low-light scenarios and real-time performance on resource-constrained devices. The natural class imbalance, with open eyes representing 85.4 % and yawning 5.2 % of samples, reflects realistic drowsiness occurrence patterns. The dataset includes comprehensive metadata, demographic information, and detailed annotation guidelines, making it suitable for training and evaluating lightweight machine learning models for automotive safety applications. This comprehensive dataset enables the development of robust drowsiness detection systems deployable on resource-constrained devices, particularly beneficial for emerging markets where cost-effective solutions are crucial. This contribution addresses the significant gap in publicly available drowsiness detection datasets optimized for edge computing platforms, particularly focusing on the practical challenges of varying illumination conditions in real-world driving scenarios. In essence, the Annotated Drowsiness Detection Dataset stands as a valuable resource for advancing real-time drowsiness detection technologies on edge computing platforms, supporting both static image classification and temporal sequence analysis approaches.