Drowsiness Detection of Construction Workers: Accident Prevention Leveraging Yolov8 Deep Learning and Computer Vision Techniques

Adetayo Olugbenga Onososen, Innocent Musonda, Damilola Onatayo, Abdullahi Babatunde Saka, Samuel Adeniyi Adekunle, Eniola Onatayo

Construction projects’ unsatisfactory performance has been linked to factors influencing individuals’ well-being and mental alertness on projects. Drowsiness is a significant indicator of sleep deprivation and fatigue, so being able to identify the cognitive and physical preparedness of workers on site to engage in construction tasks is important. As a consequence of the strenuous nature of the work involved in construction, long work hours, and environmental conditions, drowsiness is commonplace and has received less attention despite being a leading cause of accidents occurring on-site. Detecting drowsiness is essential for determining the safety and well-being of site workers. This study presents a vision-based approach using an improved version of the You Only Look Once (YOLOv8) algorithm for real-time drowsiness exposure among construction workers. The proposed method leverages computer vision techniques to analyze facial and eye features, enabling the early detection of signs of drowsiness, effectively preventing accidents, and enhancing on-site safety. The model showed significant precision and efficiency in detecting drowsiness from the given dataset, accomplishing a drowsiness class with a mean average precision (mAP) of 92%. However, it also exhibited difficulties handling imbalanced classes, particularly the underrepresented ‘Awake with PPE’ class, which was detected with high precision but comparatively lower recall and mAP. This highlighted the necessity of balanced datasets for optimal deep learning performance. The YOLOv8 model’s average mAP of 78% in drowsiness detection compared favorably with other studies employing different methodologies. The system improves productivity and reduces costs by preventing accidents and enhancing worker safety. However, limitations, such as sensitivity to lighting conditions and occlusions, must be addressed in future iterations.