Research On Dynamic Gesture Recognition Based On Millimeter Wave Radar Using Deep Learning

As a non-contact human-computer interaction method,the gesture recognition of millimeter-wave radar based on deep learning has broad application prospects in the fields of intelligent driving,telemedicine,and somatosensory games.The traditional method converts the gesture echo signal into a spectral feature map through time-frequency analysis,and then recognizes the gesture according to the method of image classification.The gesture samples obtained by preprocessing radar echoes contain less feature information,resulting in relatively poor robustness of the classification model on other data.This thesis focuses on the time correlation characteristics of millimeter wave radar echoes and the time domain modeling method based on deep learning.The main work is as follows:1.The traditional millimeter-wave radar dynamic gesture recognition method needs to perform time-frequency analysis on the distance,speed,angle and other information contained in the dynamic gesture echo,so as to obtain various characteristic spectra(sequences),and then use various classification models to These pictures or sequences are identified,and the echo processing flow is relatively cumbersome.This paper proposes a dynamic gesture recognition method for millimeter-wave radar based on One-dimensional Series connection Neural Networks(1D-Sc NN).First,the echo sequences of dynamic gestures collected from the sensor are recombined in a time-series manner,and then a one-dimensional convolution layer,a pooled sampling layer,and one-dimensional Inception v3 modules are built to obtain the feature sequences of multi-frame gestures.Finally,the features obtained by the one-dimensional convolution model are input into the Long Short-Term Memory(LSTM)network layer to aggregate the one-dimensional features in the sequence,and the temporal correlation of multi-frame sequences is used to improve the convergence speed and classification accuracy of the model.The experiment results show that the method is suitable for scenes with insufficient samples,and the accuracy of gesture classification can reach more than 96.0%,which is better than the traditional dynamic gesture recognition method based on spectral feature map.2.Traditional gesture recognition methods based on radar sensors have poor robustness in application scenarios with random dynamic interference,and are prone to misrecognition.To address this issue,a robust hand gesture recognition method is proposed based on the self-attention time-series neural networks(Atten-Ts NN).Firstly,the original radar echo is constructed in terms of frame,sequence and channel at the input terminal of the network.A one-dimensional time-series neural network is built,and the time-distributed layer is used as the wrapper to extract the feature from each frame sequence independently.Then the self-attention mechanism is employed to assign the adequate weights to the sequence of frames entered in par-allel to obtain the inter-frame correlation and suppress the random interference.Finally,the Global Avg Pooling layer is used to reduce the number of channels,and the fully connected layer outputs the label of the gesture.The experimental results show that the proposed method can achieve a high recognition rate in the presence of 25%dynamic random interference.3.When a deep learning-based dynamic gesture recognition model is deployed on an embedded device,there are problems such as a large amount of model parameters,redundant model operators and a large amount of computation,which often need to be processed by accelerated algorithms such as model pruning and compression.This paper proposes a compression method for millimeter-wave radar dynamic gesture recognition model based on knowledge distillation algorithm.First,the model structure is simplified on the basis of the Atten-Ts NN model,and the main modules are retained.Then,the model is distilled by regulating the temperature value and the teacher model(Atten-Ts NN)to guide the training method of the student model.The experimental results show that the distilled student model is greatly compressed and can still maintain an accuracy of more than 98% in the dataset with interference.