| 引用本文: | 张雨楠， 余鸿文， 石 野， 谭冠南， 盛志超， 方 勇. 基于转导条件神经自适应网络的少样本雷达手势识别[J]. 雷达科学与技术, 2025, 23(4): 387-395.[点击复制] |
| ZHANG Yunan, YU Hongwen, SHI Ye, TAN Guannan, SHENG Zhichao, FANG Yong. Few⁃Shot Hand Gesture Recognition with Radar Based on TransductiveConditional Neural Adaptive Network[J]. Radar Science and Technology, 2025, 23(4): 387-395.[点击复制] |
|
| 摘要: |
| 雷达手势识别（Hand Gesture Recognition, HGR）技术通常依赖于大量的样本数据以实现高识别精度，但是雷达数据的采集、标注需要耗费大量的人力成本。在标注数据稀缺的情况下，深度学习在HGR技术上的应用和效果均受到限制。为了解决上述问题，本文提出了一种基于转导条件神经自适应网络的手势识别算法。具体而言，该算法采用了元学习的思想，旨在通过少量训练样本训练出能够快速适应新HGR任务的模型。此外，所提算法采用了基于条件神经自适应的特征提取模块，通过调整模型中的少量参数，选择性地关注跨任务的有用特征，从而实现灵活性和鲁棒性之间的有效平衡；同时，该算法还结合了转导推理方法，利用图构造模块在数据中开发新类空间的流形结构，提高了识别精度。实验结果表明，本文提出的模型在5?way 3?shot和8?way 3?shot的少样本分类任务中，手势识别精度分别能达到90.94%和87.67%。 |
| 关键词: 手势识别 少样本学习 元学习 调频连续波雷达 |
| DOI：DOI:10.3969/j.issn.1672-2337.2025.04.004 |
| 分类号:TN958.94 |
| 基金项目:上海扬帆专项（No.23YF1412700） |
|
| Few⁃Shot Hand Gesture Recognition with Radar Based on TransductiveConditional Neural Adaptive Network |
|
ZHANG Yunan, YU Hongwen, SHI Ye, TAN Guannan, SHENG Zhichao, FANG Yong
|
|
1. School of Communication and Information Engineering, Shanghai University, Shanghai 200444, China;2. School of Information Science and Technology, ShanghaiTech University, Shanghai 201210, China;3. Huizhou Speed Wireless Technology Co Ltd, Huizhou 516000, China
|
| Abstract: |
| The high recognition accuracy of hand gesture recognition （HGR） technology with radar relies on a substantial amount of sample data. However, collecting and labeling radar data requires significant manpower costs. When faced with a small amount of labeled data, the application and effectiveness of deep learning in HGR algorithms are limi?ted. To address the above problem, a HGR algorithm based on transductive conditional neural adaptive network is proposed in this paper. Specifically, this network adopts the concept of meta?learning, aiming to train a model that can quickly adjust to new HGR tasks with a few training samples. In addition, the proposed algorithm employs an adaptive feature extraction module, which adjusts a small number of parameters to selectively focus on useful features across tasks, achieving an effective trade?off between flexibility and robustness. Furthermore, the algorithm integrates a transductive inference method, which utilizes a graph construction module to develop manifold structures of the novel class spaces in the data, thereby improving the recognition accuracy. The experimental results indicate that our model achieves recognition accuracies of 90.94% and 87.67% in 5?way 3?shot and 8?way 3?shot few?shot classification tasks, respectively. |
| Key words: hand gesture recognition few⁃shot learning meta⁃learning FMCW radar |
