Research On Gesture Recognition Method Based On 77GHz Millimeter Wave Radar

In the development of human-computer interaction technology,gesture recognition has attracted much attention.Traditional image or video-based gesture recognition methods are susceptible to interference from light conditions,bad weather,and obstructions,and there may be a risk of user 's privacy being compromised,which greatly limit its application scenarios.However,the radar signal is immune to light and weather conditions,enabling around-the-clock operation and avoiding user 's privacy leakage.Therefore,the research on gesture recognition technology based on radar signals has important theoretical significance and application value.Compared with other band radars,the 77 GHz millimeter wave radar has higher range,speed and angle measurement accuracy,and can be applied to small-scale gesture recognition scenes.In this paper,based on the 77 GHz millimeter wave radar,the four gestures are collected which include tick,radial wave,clockwise rotation and counterclockwise rotation.Then the convolutional neural network is used for feature extraction and recognition,and satisfied results are obtained.Firstly,considering the coupling phenomenon of the radar transmit-receive antenna,that is,the huge peak appears at about 8cm from the radar which directly affects the subsequent feature extraction process.This paper breaks away from the conventional idea of truncating from the distance domain,and uses the classical modal decomposition and the improved wavelet threshol d decomposition to deal with it.After that,signal-to-noise ratio and running time of the two methods are compared.Finally,the best de-interference method is selected to remove it.Secondly,in order to make use of the effective information that radar can provide as much as possible,this paper proposes a feature extraction and construction method based on diversified feature map.That is,the time-distance spectrum,the time-velocity spectrum and the time-angle spectrum of each gesture are extracted,and the data of each single characteristic spectrum is normalized,and then the three are spliced to form a diversified characteristic spectrum.In addition,in order to facilitate the training of the late convolutional neural network,each single characteristic spectrum and the diversified characteristic spectrum are separately subjected to preprocessing operations such as graying,de-equalization,and scale normalization,so that they can be directly converted into a picture feature as the convolutional neural network input.Finally,in the classification stage,this paper designs a convolutional neural network architecture for diversified feature maps.In the experimental simulation,the network training of the three single feature maps shows that the recognition accuracy of the three feature maps is more than 80%,indicating the validity and accuracy of the feature map.The diversified feature map was tested and the recognition accuracy was as high as 98%,which was 7%-15% higher than that of the single feature map,and fully demonstrated the advantages of the diversified feature map.In contrast,the time complexity is compared with the two classical networks,which shows the advantage of the convolutional neural network architecture designed in this paper.In addition,in order to analyze the network performance of this paper,the influence of gesture size and speed on the recognition accuracy rate is also evaluated.The result is also consistent with the inference of this paper.