Research On Deep Neural Network Gesture Recognition Algorithm For FMCW Radar

The rapid development of computers and communication technologies has made unprecedented changes in human society.Human Computer Interaction(HCI)has gradually shifted from computer-centric to human-centric.Traditional gesture recognition methods have been gradually replaced by Frequency Modulated Continuous Wave(FMCW)methods due to inconvenience to wear and privacy.The research contents of this thesis are as follows:1.Construct a multi-dimensional gesture parameter feature map based on FMCW radar.First,by analyzing the intermediate frequency(IF)signal of the FMCW radar,a Fast Fourier Transformation(FFT)algorithm is used to obtain the distance information of the gesture.Then,a 2D Fast Fourier Transform(2D-FFT)algorithm is used to parse the Doppler information of the gesture.After that,a multiple signal classification(MUSIC)algorithm is used to analyze the angle information of the gesture.Finally,the distance-time map(RTM),Doppler-time map(DTM),and angle-time map(ATM)of the gesture are obtained through the accumulation of multiple frames.2.Expand the gesture dataset according to the characteristics of gestures in the parameter map.First,the adversarial generation network and the multi-dimensional-single-dimensional(Mixup Augmentation,MA)method are used to augment the gesture dataset.Then,the constructed dataset is verified by the model.The results show that the accuracy of the gesture recognition by the WN dataset reaches to 94.98%.In addition,the training of adversarial generation network model is unstable,the time cost is large,and the efficiency is low.When using the WMA dataset for gesture recognition,the support for training distribution is expanded by extracting additional virtual samples from the neighborhood,and the sample size is expanded without increasing model complexity.Besides,the recognition effect reaches to 96.83%.3.Build a gesture recognition model of 3D parameter feature fusion.First,a multi-dimensional parameter map of gesture features is constructed with data augmentation methods.At the same time,a network proposed in this thesis that can effectively utilize multiple dimensional gesture feature fusion(CMFF)is composed of Refine Net and MMFNet.CMFF network can effectively fuse three-dimensional gesture features by residual learning.By connecting the complementary modal features,the feature information contained in the three images of RTM,DTM,and ATM is effectively extracted.The simplified model CMFF-LW + WMA reduces the time complexity by 2 times and reduces the space complexity by 2 times.