Research On Dynamic Gesture Recognition Method Based On FMCW Radar

Gesture recognition technology is an effective way of natural communication between human and machine.Among all kinds of recognition devices,radar has attracted more and more attention because of its strong target detection ability,insensitivity to light,background and other factors,low energy consumption and privacy protection.At present,the research of dynamic gesture recognition based on radar signal has some difficulties,such as insufficient parameter information,large amount of data processing and low recognition accuracy.In view of the above problems,a dynamic gesture recognition method based on Frequency Modulated Continuous Wave(FMCW)radar is studied by combining radar signal processing method and deep learning,which is excellent in classification and recognition.And in view of the shortcomings of existing radar recognition methods,corresponding improvements are made.The main works is as follows:(1)Aiming at the problem that the inadequate information of gesture parameters in existing radar gesture recognition methods affects the recognition accuracy and too large radar time-frequency image increases the complexity of recognition model.A method of constructing multi-dimensional gesture parameter vector is proposed.The parameters of distance,velocity and azimuth of gesture are extracted by distance,velocity and space Fast Fourier Transformation.Solving the problem of inaccurate azimuth detection caused by too few radar antennas,high-order cumulants and Fast Fourier Transformation are combined to realize angle estimation in the extraction of azimuth angle parameters of gesture.In the construction of gesture parameter vector,target focusing is combined with CFAR automatic target detection algorithm to improve detection performance.The multi-parameter vector matrix composed of X-Y coordinates and velocity vectors of the detected gesture target in the radar plane is used as the input of the neural network for gesture recognition.This method can effectively avoid the clutter noise and the image size too large when using the time-frequency map of gesture parameters,and combine the speed,distance and angle parameters to enrich the gesture recognition information.The experimental results show that the recognition rate of gesture recognition method using multi-dimensional parameter information is 2.05% and 1.19% higher than that using only one-dimensional velocity information or two-dimensional distance and velocity information.In addition,compared with other methods in the literature,it can effectively reduce the complexity of the neural network model.(2)Aiming at the problem of insufficient recognition rate in current research of radar gesture recognition,two models for gesture classification are designed,namely,Convolution Neural Network(CNN)and Long Short Term Memory Network(LSTM).Combining the advantages of the two models,a hybrid model of CNN and LSTM is designed.The model first extracts the short-term features of gestures through CNN network,then learns the features through LSTM network,and finally classifies gestures.The experimental results show that the recognition accuracy of the hybrid model is improved by 4.02% and 1.59% respectively compared with the CNN model and the LSTM model.(3)On the basis of the above three studies,a complete solution of dynamic gesture recognition using FMCW radar signal target detection capability is proposed.A gesture-controlled music player system is developed.Through four kinds of gesture control music player to start playing music,stop playing music,switch the previous song and switch the next song these four functions,verify the feasibility of this scheme.