| Popis: |
Personal Protective Equipment (PPE) plays a crucial role in protecting workers from unpredictable physical threats and reducing fatal occupational injury rates. However, PPE is not always worn by workers in all cases, which poses a challenge to site management. In this context, object detection methods based on deep learning have been used to strengthen site management. To address the challenges posed by limited PPE detection performance under complex environmental conditions, we develop the MARA-YOLO object detection model, which is designed to balance speed and accuracy. Firstly, based on YOLOv8-s, we introduce a modified re-parameterizable backbone, which consists of MobileOne-S0 and an efficient downsampling block known as the Attentional Space-to-Depth Block (AS-Block). Subsequently, we propose the R-C2F module, which fused feature maps from diverse receptive fields and enhances the model’s sensitivity to the texture information of objects, as well as its ability to capture information from varying depths. Build upon R-C2F and Adaptively Spatial Feature Fusion (ASFF), a multi-scale feature fusion module RASFF is further introduced to mitigate inconsistent multi-scale outputs in the model. Finally, a dedicated dataset consisting of 2750 images covering 9 categories is constructed. The ablation experiments demonstrate that compared to the baseline, MARA-YOLO achieves a 6.7% improvement in AP50 and a 10.2% improvement in AP75 on the proposed KSE-PPE dataset. In the comparative experiments, MARA-YOLO achieves a mean average precision (mAP) of 74.7% on the KSE-PPE dataset, surpassing other lightweight state-of-the-art models by more than 4.95%. |
