融合注意力机制的毫米波雷达人体动作识别方法

Human motion recognition method of millimeter-wave radar integrated with attention mechanism

为解决少样本场景下毫米波雷达人体动作识别过程中卷积神经网络(CNN)易出现过拟合、训练效果不理想等问题,提出一种融入时序注意力机制的CNN和视觉转换器模型结合的方法.该方法首先对收到的雷达回波信息做预处理,再通过短时傅里叶变换(STFT)进行时频分析得到时频图,最终将带有特征信息的图像送入融合的网络模型中进行分类识别.实验结果表明,与其他4种模型的方法相比,本文提出的方法识别准确率最高,识别效果可达到91.57%.该方法能有效地增强网络对于时间维度建模,增加了网络收敛速度,达到了提升识别准确率的效果.

In order to solve the problems that convolutional neural networks are prone to overfitting and unsatisfactory training results in the process of human motion recognition by millimeter wave radar in the scenarios with few samples,this paper proposes an approach combining convolutional neural network with the vision transformer(VIT)model that incorporates a temporal attention mechanism.Firstly,this method is used to preprocess the received radar echo information.Then the time-frequency graph is obtained through the time-frequency analysis by short-time Fourier transform(STFT).Finally,the images with characteristic information are fed into the integrated network model for classification and recognition.The experimental results show that compared with the other four models,the proposed method has the highest recognition accuracy,with the recognition effect reaching 91.57%.This method can effectively enhance the time dimension modeling of the network,increase the convergence speed of the network,and achieve the effect of improving recognition accuracy.