Adaptive Feature Concatenation in YOLOv8n for Construction Site Safety Detection

Monitoring Personal Protective Equipment (PPE) compliance is critical for construction safety, yet traditional lightweight detectors often struggle with severe occlusion and scale variations due to static feature fusion mechanisms. The main research contribution is the development of an enhanced YOLOv8n architecture integrating a novel Adaptive Feature Concatenation Module (AFCM) to address these limitations. Unlike standard concatenation, the proposed method employs AFCM in the neck network to dynamically recalibrate input features using learnable scalar weights normalized via Softmax. This mechanism allows the model to selectively emphasize semantically rich features while suppressing background noise without increasing channel dimensions. Experimental validation on the Construction Site Safety Image Dataset (CSSID) demonstrates that the model achieves an mAP@50:95 of 51.42%, surpassing the baseline YOLOv8n by 3.45%. Comparative analysis confirms that the proposed method outperforms state-of-the-art lightweight detectors, including YOLO-GP and MKD-YOLO, particularly in recovering small and occluded targets. Crucially, these improvements are achieved while maintaining 3.00 million parameters and 8.1 GFLOPs, identical to the baseline, ensuring no additional computational overhead. Consequently, the proposed framework offers a viable and effective solution for real-time, automated safety monitoring in resource-constrained edge devices.