YOLOv8n-ASF-DH: An Enhanced Safety Helmet Detection Method

Abstract

Wearing a safety helmet is an essential aspect of safety production management. To address the issue of low detection accuracy of existing safety helmet detection algorithms for small targets and complex scenes, we propose the YOLOv8n-ASF-DH model (DH stands for Dynamic Head), which is based on YOLOv8n and integrates the Attentional Scale Sequence Fusion (ASF) structure and the Dynamic Head (DyHead). In the backbone layer, a Triplet Attention mechanism is incorporated to enhance the model’s focus on features of small targets. In the neck layer, the ASF structure efficiently fuses different level output features extracted by the backbone network, enhancing the overall feature representation capability of the model. In the detection head, DyHead adjusts the relationships between different feature layers, enhancing the model’s performance on different scales and complex scenes. Additionally, adopting the Focal-EIoU loss function balances the contributions of high-quality and low-quality samples in loss calculation. Experimental results demonstrate that the improved model enhances detection performance in complex scenes and small target scenarios. Compared to the YOLOv8n model, the YOLOv8n-ASF-DH model achieved an improvement of 3.066% in accuracy and 3.883% in recall. Additionally, the model exhibited an increase of 2.584% in mAP 0.5 and 6.131% in mAP 0.5:0.95. Compared with other mainstream object detection algorithms, the improved model significantly enhances the performance of safety helmet target detection tasks.

Keywords

• Helmet wearing detection
• YOLOv8
• attentional scale sequence fusion
• dynamic head
• triplet attention mechanism