<p>Due to the limited storage and computational capabilities of equipment at construction sites and the complex on-site conditions that can lead to target misdetections and omissions, it is necessary to minimize computational overhead and storage usage as much as possible while maintaining detection performance. This paper proposes an enhanced safety helmet detection method based on the optimized YOLO11 model, addressing two key challenges: detection accuracy in complex backgrounds and model deployment on edge devices. We introduced PConv in the C3K2 module to reduce computational costs and improve feature extraction efficiency, providing high-quality input for multi-scale processing. SPP-ELAN integrates multi-scale information extraction with feature aggregation, capturing contextual information across scales and enhancing local features with global context, offering optimal input for ECA. ECA optimizes the channel weight distribution of SPP-ELAN output features, capturing channel dependencies, highlighting important features, suppressing background noise, and enhancing multi-scale feature expression and sensitivity to small targets. In addition, we propose the NWD–CIoU loss function, combining NWD and CIoU to inherit CIoU’s positioning advantages and address its limitations in non-overlapping targets and scale variations, improving small target localization and multi-scale detection for a more robust and efficient model. Experimental results show that our improved model achieves a 1.7% increase in recall, a 24.5% reduction in parameter count, and a 6.7% improvement in detection speed while maintaining precision and mAP metrics. These improvements enhance the model’s performance in complex environments and small target detection, providing a viable solution for practical applications in safety helmet real-time monitoring systems and edge device deployment.</p>