结合GhostNetv2与YOLOv7的交通标志实时检测
Real-time detection of traffic signs combining GhostNetv2 and YOLOv7
黄伟涛 1许钡榛 2黄茂荣3
作者信息
- 1. 广州市城市更新规划设计研究院有限公司,广东 广州 510031
- 2. 广州市城市规划勘测设计研究院有限公司,广东 广州 510053
- 3. 广州中鹤建设有限公司,广东 广州 511400
- 折叠
摘要
针对现有轻型检测方法在交通标志检测中检出率低且鲁棒性差的问题,提出一种改进YOLOv7的轻型交通标志检测算法.在骨干网络中,引入第二代幽灵卷积核替代现有卷积核结构.在特征融合网络中,让来自骨干网络的更大尺寸的特征图参与信息融合.在训练阶段,使用带有焦点机制的高效交并比函数计算目标框回归损失,以协调梯度函数计算目标分类损失.在测试阶段,通过TensorRT对模型轻量化部署.实验结果表明,改进模型精度较YOLOv7提高9.29%,检测速度达到42.16 m·s-1,适合部署在低功耗硬件中开展实时交通标志检测任务.
Abstract
In view of the low detection rate and poor robustness of existing lightweight detection methods for traffic sign detection,an improved YOLOv7 lightweight traffic sign detection algorithm was proposed.In the backbone network,a second-generation ghost convolution kernel was introduced to replace the existing convolution kernel structure.In the feature fusion network,larger-sized feature maps from the backbone network were involved in information fusion.In the training stage,an efficient intersection over union ratio function with a focus mechanism was used to calculate the target box regression loss,so as to coordinate the gradient function to calculate the target classification loss.In the testing stage,the model was deployed in a lightweight manner through TensorRT.The experimental results show that the improved model accuracy is 9.29%higher than YOLOv7,and the detection speed reaches 42.16 m·s-1,which is suitable for deployment in low-power hardware for real-time traffic sign detection tasks.
关键词
交通标志检测/轻量级网络/第二代幽灵卷积/解耦全连接注意力(DFC)/改进损失函数Key words
traffic sign detection/lightweight network/second-generation ghost convolution/decoupled full connected attention(DFC)/improved loss function引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据