Research On Gesture Recognition Technology Based On Range Doppler Information Extraction And Deep Learning

The gesture recognition technology of millimeter wave radar can effectively make up for the limitations of gesture recognition technology based on optical sensors or wearable devices by virtue of its advantages such as strong environmental adaptability,privacy protection and easy to realize miniaturized applications.In recent years,it has become a research hotspot in the field of radar sensor intelligent application technology.At present,millimeter wave radar gesture recognition technology mainly extracts features based on the range,Doppler and angle of gestures,and realizes gesture recognition by multi feature fusion.However,for gestures with insignificant changes in range and angle,there are problems of feature fusion failure and unable to accurately recognize small amplitude and small angle motion gestures.Compared with the classification methods,most of them use machine learning algorithms.It uses gesture feature map to realize gesture recognition,but the performance of the model is poor under the condition of small samples.To solve the above problems,this paper studies the gesture recognition technology based on gesture range Doppler image feature extraction and deep learning algorithm.The research content has theoretical significance and important application value for radar sensor gesture recognition.The research work of this paper includes:Firstly,the system parameter configuration,data preprocessing and feature extraction methods of millimeter wave radar gesture are studied.For the problem that the feature representation of small amplitude and small angle motion gesture is not obvious from range and angle,the hardware performance of radar system and the method of gesture motion feature extraction are studied.Then,the data acquisition and preprocessing process of 77 GHz millimeter wave array radar sensor are studied,the feature extraction methods of three feature maps of small amplitude motion gesture based on range Doppler image are proposed,and eight common gesture data sets are constructed.Secondly,gesture recognition technology based on single branch convolutional neural network is studied.Aiming at the poor performance of deep convolution neural network under the condition of small samples,the factors affecting the model performance and solutions are studied,and an improved training method based on model migration is proposed.The experimental results show that the improved training method based on model migration can effectively alleviate the problems of difficult training and over fitting of deep neural network model under the condition of small samples,and improve the performance of the network model.Compared with conventional training methods,the improved training method improves the gesture recognition accuracy by more than 10% on average.Among them,the improved training method based on model migration can achieve 97.7% gesture recognition accuracy on resnet34 network.Finally,gesture recognition technology based on multi branch depth convolution neural network is studied.Aiming at the problem of insufficient representation and missing features of single feature on gesture motion features,the multi branch convolution neural network model is further built to fuse the three feature maps of small amplitude motion gestures.The experimental results show that compared with the single branch model,the multi branch model effectively improves the performance of gesture recognition through multi feature fusion,The recognition accuracy of 8 common gestures is improved by 5% and the effectiveness of the feature extraction method in this paper is further verified.The training method based on transfer learning proposed in this paper effectively improves the recognition accuracy under the condition of small samples.The gesture recognition method based on multi branch deep convolution neural network model improves the performance of the model through multi feature fusion,which is of great significance to improve the accuracy of millimeter wave radar gesture recognition.