首页|基于RD-YOLO的毫米波雷达和视觉融合显著性目标检测

基于RD-YOLO的毫米波雷达和视觉融合显著性目标检测

扫码查看
原文链接
  • NETL
  • NSTL
  • 万方数据
为了解决单传感器在复杂环境下目标检测精度低问题,提出了一种基于RD-YOLO的毫米波雷达和视觉融合的显著性目标检测方法.首先设计了能够将毫米波雷达点云转换为图像的方法,使毫米波雷达和视觉数据在模型输入时实现特征融合;然后通过动态互补注意力机制,对两个图像分支生成特征设置空间和通道动态注意力权重;最后采用YOLOv8 检测融合后特征,引入改进损失函数Focal Loss以解决样本不均衡问题.在数据集nuScenes上开展的相关实验表明,与YOLOv5、YOLOv7、YOLOv8、Faster R-CNN和FCOS相比,所提方法目标检测综合性能良好,均值平均精度比原始YOLOv8 提升了9.19%.
Salient target detection using millimeter wave radar and visual fusion based on the RD-YOLO
To address the problem of low detection accuracy in single sensor target detection in complex environments,this paper proposes a salient target detection method based on the RD-YOLO using millimeter wave radar and visual fusion.Firstly,a method was designed to convert millimeter wave radar point clouds into images,enabling feature fusion between millimeter wave radar and visual data during model input.Then,the spatial and channel dynamic attention weights were set for the two image branching generation features through an interactive complementary attention mechanism.Finally,using YOLOv8 to detect the fused features,an improved loss function Focal Loss was introduced to solve the problem of imbalanced samples.Relevant experiments were conducted on the nuScenes dataset.The results show that,compared with YOLOv5,YOLOv7,YOLOv8,Faster R-CNN,and FCOS,the method proposed in this paper has good overall performance in object detection,with an average accuracy improvement of 9.19% compared to the original YOLOv8.

salient target detectionfeature fusionmillimeter-wave radarradar point cloud to imagedynamic complementary attention module

王杨、王臣飞、张广海、张俊、后海伦、欧阳少雄

展开 >

芜湖学院 大数据与人工智能系,安徽 芜湖,241000

安徽师范大学 计算机与信息学院,安徽 芜湖,241002

芜湖市大数据与人工智能工程技术研究中心,安徽 芜湖,241003

显著性目标检测 特征融合 毫米波雷达 雷达点云转换 动态互补注意力机制

安徽省高等学校自然科学研究重点项目安徽省高等学校优秀青年人才支持计划安徽省高等学校学科(专业)拔尖人才学术资助项目

2022AH052899gxyq2022167gxbjZD2022147

2024

邵阳学院学报(自然科学版)
邵阳学院

邵阳学院学报(自然科学版)

影响因子：0.286
ISSN：1672-7010
年,卷(期)：2024.21(4)
  • 2