Performance Evaluation of YOLOv8-Pose for Vision-Based Fatigue Detection via Multi-Feature Behavioral Analysis

Abstract

Driver fatigue is a significant contributor to road traffic accidents, motivating the development of non-invasive vision-based monitoring systems. This study evaluates a single-stage YOLOv8n-Pose approach for fatigue detection based on multiple behavioral indicators, including eye closure, a hand-at-mouth gesture as a yawning-related cue, and head tilt. Unlike conventional multi-stage approaches that require separate face detection and facial landmark extraction, the proposed approach jointly detects the facial region and estimates eight task-specific keypoints within a single inference process. The model was trained on 2,625 images representing variations in lighting conditions, subjects, accessories, and camera distance. Performance was evaluated using bounding-box and keypoint detection metrics, followed by live testing under different head orientations and camera distances. The model achieved a bounding-box mAP50 of 0.967 and a keypoint mAP50 of 0.817, with F1-scores of 0.90 for Eyes Open, 0.94 for Eyes Closed, and 0.97 for Hand-at-Mouth detection. Head-angle testing showed a monotonic decrease in ear-to-shoulder keypoint distance as the head tilted from the upright position, supporting its use as a head-tilt indicator within the tested conditions. Live testing identified an effective camera distance of 60–100 cm and an average processing rate of approximately 3.3–3.5 FPS. Error analysis identified backlighting as a source of incorrect eye-state predictions and hand occlusion as a cause of missed Hand-at-Mouth detections. These results demonstrate the potential of a single-stage pose-based approach for multi-indicator vision-based fatigue detection while highlighting the need for improved lighting robustness, diverse hand-pose training data, and higher inference efficiency.