基于毫米波雷达的手势识别在体感游戏、远程操控、医疗监护等领域具有广阔的应用前景,现有的研究大多利用卷积神经网络、循环神经网络等对雷达提取到的二维手势特征图进行处理,以实现手势分类。对于复杂的动态手势,存在训练效率低、并行能力差等不足。针对上述问题,文中提出了一种基于距离-多普勒-角度多维特征的卷积自注意力机制网络(CVTNet)模型。CVTNet 模型针对射频手势信号,先利用卷积层对距离-多普勒-角度图进行特征提取,再使用Vision Transformer 进行手势的识别与分类。仿真结果表明,基于CVTNet 的手势识别算法拥有更高的识别准确率,达到96. 29%。同时,该算法在微手势及静态手势的识别上具有更大的优势。

Millimeter wave radar-based gesture recognition has broad application prospects in such fields as physical games, remote control, and medical monitoring. Most of the existing researches use convolutional neural network or recurrent neural network to process the two-dimensional gesture feature map extracted from radar to achieve gesture classification. However, they are inefficient in handling complex dynamic gestures and have poor parallelism. To solve these problems, a gesture recognition network named convolution and vision transformer network (CVTNet) based on multi-dimensional features of distance-Doppler-angle is proposed. CVTNet first uses the convolutional layer to extract features from the range-Doppler-angle map of RF gesture signal, and then uses Vision Transformer to recognize and classify gestures. Simulation results show that, the gesture recognition algorithm based on CVTNet has a higher recognition accuracy of 96. 29%. Meanwhile, it has better performance in the recognition of micro gestures and static gestures.

TP391.4

国家自然科学基金资助项目(62272242, 61902237); 江苏省研究生科研与实践创新计划资助项目(KYCX22_0986)