| 摘 要: 为解决工地场景下安全帽检测因背景复杂、目标密集等导致YOLOv8n算法识别精度低、易出现漏检及误检等问题,提出一种改进YOLOv8n的RDCA-YOLO安全帽检测算法。通过改进Backbone结构提升模型特征提取能力;嵌入CBAM注意力机制,增强对小目标信息提取的能力;设计一种Coord-BiFPN结构,增强网络特征融合能力;提出一种OD-C2f结构,实现提取不同形状和大小安全帽的关键特征;设计一种FR-DyHead检测头替换原始Detect结构,提升了检测精度并降低了延时。使用GDUT-HWD数据集进行训练测试,结果表明所提算法的mAP 达到85.8%,相比YOLOv8n提升了2.6%,能有效提高复杂场景下的安全帽佩戴检测精度。 |
| 关键词: YOLOv8;CBAM注意力机制;Coord-BiFPN;FR-DyHead检测头 |
|
中图分类号: TP391.4
文献标识码: A
|
| 基金项目: 科技创新2030———“新一代人工智能”重大项目(2018AAA0100400);湖南省自然科学基金项目(2021JJ50058);湖南省自然科学基金(2022JJ50051) |
|
| Helmet Wearing Detection Algorithm in Complex Scenes Based on Improved YOLOv8n |
|
LEI Yuanyi, ZHU Wenqiu, LIAO Huan
|
(School of Computer, Hunan University of Technology, Zhuzhou 412007, China)
yuanyi_lei@foxmail.com; wenqiu_zhu@126.com; huanliao@foxmail.com
|
| Abstract: Complex background and dense targets of helmet detection in construction site scenes lead to low recognition accuracy, easily missed detection and false detection of YOLOv8n algorithm. To solve these problems, this paper proposes a RDCA-YOLO safety helmet detection algorithm based on improved YOLOv8n. Firstly, model feature extraction ability is enhanced by improving Backbone structure and CBAM (Convolutional Block Attention Module) attention mechanism is embedded to improve small target information extraction. Secondly, a Coord-BiFPN structure is designed to enhance network feature fusion capability, and an OD-C2f structure is proposed to extract key features of safety helmets with different shapes and sizes. Finally, a FR-DyHead detection head is designed to replace the original Detect structure, so as to improve detection accuracy and reduce latency. Training and testing are conducted on GDUTHWD helmet dataset. Test results show that the mAP of the proposed algorithm achieves 85.8% accuracy, which is 2.6% higher than YOLOv8n, which verifies that the proposed algorithm can effectively improve the accuracy of helmet wearing detection in complex scenes. |
| Keywords: YOLOv8; CBAM attention mechanism; Coord-BiFPN; FR-DyHead detection head |
