| 摘 要: 针对当前电动车头盔佩戴检测在密集场景下误检率高、小目标检测的漏检率高、模型体积大等问题,提出了改进的YOLOv8头盔佩戴检测方法。首先,使用轻量化卷积模块替换主干网络中的常规卷积,降低网络的参数量。其次,增加一个检测层,用来提高网络对小目标的检测能力;在Neck部分,使用基于DCNv3改进的C2f_DCN模块,并融入EMA注意力机制。最后,采用Inner-MPDIoU(Inner Multi-Point Distance Intersection over Union)损失函数代替了原来的损失函数,用于改善边界框纵横比的收敛速度。实验结果显示,改进后的模型在公开数据集上的mAP@0.5达到了88.9%,相较于基准模型YOLOv8提高了5.4%,并且模型的体积压缩至原来的36.6%,更加适用于实际交通场景中的部署需求与应用场景。 |
| 关键词: YOLOv8 头盔检测 注意力机制 轻量化 |
|
中图分类号: TP391
文献标识码: A
|
| 基金项目: 陕西省自然科学基础研究计划(2022JQ-175),陕西省教育厅专项科研计划(22JK0303),陕西科技大学科研启动项目(2020BJ-18) |
|
| Helmet Wearing Detection Method Based on YOLOv8 |
|
ZHANG Yanqiang, YAO Bin, WANG Meijia
|
(School of Electronic Information and Artificial Intelligence, Shaanxi University of Science & Technology, Xi’an, 710021, China)
z13720917730@163.com; yaobin@sust.edu.cn; 4672@sust.edu.cn
|
| Abstract: To address issues such as high false detection rates in dense scenarios, high miss rates for smal-l target detection, and large model size in current electric vehicle helmet detection, an improved YOLOv8-based helmet detection method is proposed. Firstly, lightweight convolution modules replace standard convolutions in the backbone network to reduce parameters. Secondly, an additional detection layer is introduced to enhance smal-l target detection capability. In the neck section, a C2f_DCN module improved with DCNv3 is adopted, incorporating the EMA attention mechanism. Finally, the Inne-r MPDIoU loss function replaces the original loss function to accelerate the convergence of bounding box aspect ratios. Experimental results show that the improved model achieved an mAP@0.5 of 88.9% on public datasets, a 5.4% increase over YOLOv8. The model size is compressed to 36.6% of the original, making it more suitable for deployment in rea-l world traffic scenarios. |
| Keywords: YOLOv8 helmet detection attention mechanism lightweight |
