Research On Millimeter Wave Radar Gesture Recognition Algorithm Based On Small Samples

In the field of gesture recognition,compared with data gloves,cameras and other sensors,radar has the advantages of not being affected by ambient light,not exposing user privacy,and less data.At the same time,compared with traditional long band radar,millimeter wave radar is more suitable for gesture recognition tasks with better resolution and smaller volume.In the latest research on gesture recognition related algorithms,thanks to the advantages of millimeter wave radar,more and more researchers choose millimeter wave radar as the sensor to achieve gesture classification recognition,which proves that gesture recognition algorithm based on millimeter wave radar has important research significance.However,the current algorithms for gesture recognition based on millimeter wave radar have problems such as difficult data acquisition,complex classification model and too large dimension of input data.To solve the above problems,this paper uses fixed category feature extraction instead of deep learning network to extract features,uses the generated countermeasure network to generate radar 2DFFT matrix data,and uses the self attention mechanism network instead of the combined network of convolutional neural network and cyclic neural network to achieve gesture recognition classification.Through network training and testing,this method is proved to be feasible.The main research content is as follows:1.The paper briefly describes the millimeter wave radar system and analyzes the resolution of the millimeter wave radar;Methods for obtaining distance,velocity,and angle information;Engineering Implementation Method(3D-FFT);Static clutter interference suppression;In this paper,millimeter wave radar parameter configuration for gesture recognition task.At the same time,the environment and guidelines for collecting gesture data in this article were introduced,as well as the hardware board devices required for collecting gesture echo data in this article.2.In response to the complex problem of millimeter wave radar data acquisition,this paper proposes an RF-GAN(Radar Feature GAN)network based on generative adversarial networks and a combination network RF-LSGAN(Radar Feature LSTM-GAN)using generative adversarial networks to extend short-term memory networks for generating gesture radar echo data.These two types of networks were trained separately,and the generated data was evaluated in terms of single frame 2D-FFT heatmap and continuous frame feature changes over time.Based on the training results,the RF-GAN network was selected to generate a large amount of gesture radar echo data in this paper.3.To address the complexity of gesture classification networks,an RFT(Radar Feature Transformer)network based on self attention mechanism is proposed.Firstly,in order to correctly and effectively associate temporal information,the advantages and disadvantages of various temporal encoding methods were analyzed,and the reasons for choosing learnable onedimensional position encoding in this paper were explained.Secondly,according to the data set and gesture recognition classification tasks required in this paper,the basic framework of the RFT network is determined,and the network structure and parameters are determined through training.Finally,compared with the existing gesture recognition classification network in three dimensions of classification accuracy,number of gesture categories and recognition difficulty,it proves that RFT network performs well in multi category gesture classification.