Research On Gesture Recognition Based On Millimeter Wave Radar

Gesture recognition can be applied to intelligent car driving,motion sensing games,medical care and other application scenarios.At present,camera-based gesture recognition is relatively mature.However,this kind of method is easily affected by the change of ambient light and has great limitations in application.In this paper,millimeter-wave radio frequency technology is used for gesture recognition,which can overcome the application limitation of video image recognition,and the radar signal penetration ability is strong,and the recognition accuracy is high.It has high integration and reliability.In this paper,TI IWR1642 77 G millimeter wave radar is used to study eight kinds of dynamic gesture recognition.Firstly,the principle of gesture ranging and speed measurement of frequency modulated continuous wave millimeter-wave radar is introduced.Considering the problems caused by different gesture,environment and distance factors in data set construction,Eight kinds of dynamic gesture data are designed and processed,including left-right wave,forward-backward push and pull,turning toward radar,right-up wave,left-down wave,up-down wave,up-down wave,open and close,and clenched fist.A total of 8000 different kinds of gesture data were collected by using frequency modulated continuous wave radar experimental platform.Secondly,the gesture echo data is analyzed,the gesture data is extracted from the fixed distance cells,and the multiple distance cells where the gesture is located are denoised by pulse average coherent accumulation.Then,short-time Fourier Transform(STFT)and distance Fourier Transform(FFT)are applied to the slow-time(impulse)dimension data.The time-frequency and speed-time spectra of doppler frequency with time which can express eight kinds of dynamic gesture features can be obtained.Then,the Moving Target Indicator(MTI)algorithm is used for zero-frequency denoising of the acquired gesture data.Thirdly,to improve gesture recognition accuracy,an improved network architecture is proposed that replaces the residual blocks in RESNET18 with the Inception V1 module,leaving the residual links in the original RESNet18 network architecture unchanged.The improved network architecture enables the use of Inception V1 module with 4 branch channels at the original resnet18 residual block location,and the convolutional kernels of5×5,1×1 or 3×3 on the four branch channels are used to extract the features of gesture images in different dimensions respectively.Finally,through experimental verification,the improved network architecture reduces the number of operating parameters by 4 times and improves the running speed by 2-3times compared with the previous network architecture without improvement.The improved network architecture has an accurate recognition rate of more than 98% for each type of gesture,both in the empty frame and in the complex lab scene where people move around frequently.In terms of test accuracy and verification accuracy of test set,the improved network architecture is better than the original network architecture,which proves the feasibility of the improved network architecture.