随着智能化时代的到来,基于毫米波雷达的手势识别逐渐成为研究热点。针对目前基于雷达的手势识别方法中存在利用信息有限且泛化能力低的问题,文中在三维卷积神经网络的基础上提出了三维联合识别网络(3DURNet)模型。3DURNet 以宽带雷达获取的距离-多普勒(RD)图序列作为输入,使用金字塔注意力卷积模块提取RD 序列的多尺度空间特征并自适应地校准跨维度的通道权重,强化重要特征,进一步使用时间自注意力模块对全局时序信息进行建模,最后通过分类器得到识别结果。文中利用毫米波雷达在多种场景下对多名实验对象的不同手势动作进行测量形成一套雷达手势识别数据集。对比实验表明:所提出的3DURNet 网络模型对手势动作识别的准确率可达到95. 6%,参数量比主流3D网络降低一个数量级,同时具有良好的泛化能力,可为基于雷达技术的手势动作识别提供新的技术方案。

With the advent of the intelligent era, gesture recognition based on millimeter-wave radar has gradually become a research hot spot. To address the problems of limited utilization of information and low generalization ability in current radar-based gesture recognition methods, a three-dimensional joint recognition network (3DURNet) model based on 3D-CNN is proposed in the study. Firstly, 3DURNet takes range-Doppler (RD) map sequence acquired by broadband radar as the input. And then pyramidal attention convolution module is used to extract the multiscale spatial of RD sequence features and adaptively calibrate the channel weights across dimensions to reinforce the important features. Subsequently, the temporal self-attention module is used to model the global timing information. Finally, the recognition results are obtained using a classifier. In this paper, a radar gesture recognition dataset is formed by measuring different gesture actions of multiple experimental subjects in multiple scenarios using millimeterwave radar. Comparative experiments show that the proposed 3DURNet model can achieve an accuracy of 95. 6% for gesture action recognition with less number of parameters and good generalization ability. The proposed 3DURNet provides a new technical solution for gesture action recognition based on radar technology.

TN975.51

国家自然科学基金资助项目(61901487, 61871384,61921001); 湖南省自然科学基金资助项目(2021JJ40699); 中国博士后科学基金资助项目(2021TQ0084)