| 引用本文: | 李家强， 汪星宇， 陈金立， 姚昌华. 基于轻量化卷积神经网络车载雷达图像目标识别方法[J]. 雷达科学与技术, 2025, 23(1): 82-91.[点击复制] |
| LI Jiaqiang, WANG Xingyu, CHEN Jinli, YAO Changhua. Object Recognition Method for Automotive Radar Images Based on Lightweight Convolutional Neural Network[J]. Radar Science and Technology, 2025, 23(1): 82-91.[点击复制] |
|
| |
|
|
| 本文已被：浏览 245次 下载 110次 |
码上扫一扫！ |
|
|
| 基于轻量化卷积神经网络车载雷达图像目标识别方法 |
|
李家强， 汪星宇， 陈金立， 姚昌华
|
|
南京信息工程大学电子与信息工程学院， 江苏南京 210044
|
|
| 摘要: |
| 针对车载毫米波雷达距离?方位图像细节模糊、目标占比小，卷积神经网络模型复杂难以在端侧部署的问题，本文提出了一种基于轻量化卷积神经网络YOLOv5s的车载雷达图像目标识别方法。首先结合Ghost卷积设计轻量化解耦头，并行处理检测与分类问题；其次设计融合注意力机制的Concat_att模块并引入更具边界框定位敏感性的网络损失函数EIoU Loss，充分提取特征图中小目标细节信息，加速网络收敛，提升网络精度；最后通过Slim剪枝进一步压缩模型存储空间和计算量。实验结果表明，当模型大小缩减至原始YOLOv5s网络的76.8%时，mAP@0.5与mAP@0.5:0.95较原始网络分别提升了2.7%和2.8%，适用于小目标检测，并能同时满足目标识别精度与实时性要求，适合部署至车载嵌入式系统中。 |
| 关键词: 雷达图像 YOLOv5s 轻量化 注意力机制 模型剪枝 |
| DOI：DOI:10.3969/j.issn.1672-2337.2025.01.009 |
| 分类号:TN957.51 |
| 基金项目:国家自然科学基金（No.62071238） |
|
| Object Recognition Method for Automotive Radar Images Based on Lightweight Convolutional Neural Network |
|
LI Jiaqiang, WANG Xingyu, CHEN Jinli, YAO Changhua
|
|
School of Electronics and Information Engineering, Nanjing University of Information Science and Technology, Nanjing 210044, China
|
| Abstract: |
| To address the issues of blurry details and small target proportions in automotive millimeter?wave radar images, as well as the complexity of convolutional neural network models that are difficult to deploy on the edge, an automotive radar image target recognition method based on lightweight convolutional neural network YOLOv5s is proposed. First, a lightweight decoupled head is designed by incorporating Ghost convolution, enabling parallel processing of detection and classification tasks. Next, the Concat_att module enhanced with attention mechanism is designed, and a more boundary?sensitive network loss function EIoU Loss is introduced to fully extract detailed information of small objects in feature maps, accelerating network convergence and improving accuracy. Finally, Slim pruning is applied to further compress the storage space of the model and reduce computational complexity. The experimental results indicate that when the model size is reduced to 76.8% of the original YOLOv5s network, the mAP@0.5 and mAP@0.5:0.95 are respectively improved by 2.7% and 2.8% compared to the baseline network. This method is suitable for small target detection and meets both the precision and real?time requirements of target recognition, making it appropriate for deployment in automotive embedded systems. |
| Key words: radar images YOLOv5s lightweight attention mechanism model pruning |
|