An occlusion-aware framework for vision-based PPE detection using YOLO11x and bytetrack

Source abstract

Background: Ensuring workers' compliance with Personal Protective Equipment (PPE) remains challenging in construction environments because frequent worker movement and occlusion reduce the effectiveness of vision-based monitoring systems. Although recent studies have achieved promising object detection performance, limited attention has been given to the integrated evaluation of object detection and multi-object tracking under varying occlusion conditions, resulting in insufficient understanding of worker identity preservation during continuous monitoring. Aim: This study aims to provide an integrated evaluation of object detection and multi-object tracking for vision-based PPE monitoring by combining the YOLO11x object detection model with the ByteTrack tracking algorithm. The study investigates how different occlusion conditions influence both detection accuracy and identity preservation, providing a more comprehensive assessment of integrated PPE monitoring performance. Methods: The proposed framework was trained and evaluated using the publicly available PPE Detection Dataset, which provides annotated images of construction workers wearing different PPE categories. The framework integrates YOLO11x for object detection and ByteTrack for multi-object tracking, and its performance was evaluated under four predefined occlusion scenarios (no occlusion, partial occlusion, overlapping occlusion, and heavy occlusion) to investigate the influence of occlusion on detection and tracking performance. Result: The proposed framework achieved an overall Precision of 0.830, Recall of 0.790, mAP@50 of 0.838, and mAP@50–95 of 0.490 for PPE object detection. The highest tracking performance was obtained under the no occlusion scenario (MOTA = 0.7262, IDF1 = 0.7271, HOTA = 0.5503), whereas tracking performance progressively declined as the level of occlusion increased, with heavy occlusion producing the largest number of identity switches. These results indicate that occlusion had a greater impact on identity preservation than on object detection performance. Conclusion: The findings demonstrate the feasibility of integrating YOLO11x and ByteTrack for simultaneous PPE detection and worker tracking under varying occlusion conditions. The study further highlights that evaluating object detection together with multi-object tracking provides a more comprehensive assessment of vision-based PPE monitoring systems, particularly by revealing the greater influence of occlusion on identity preservation than on object detection performance.