Research On Micro-Motion Gesture Recognition Based On Millimeter Wave Radar And CTN Algorithm

Gesture recognition has always been a popular research topic in the field of humancomputer interaction.This is because as non-contact gesture recognition technology is increasingly used in various fields,the development prospects of this technology have become broader.In this regard,millimeter-wave radar technology has been widely used in motion target detection due to its high accuracy,ease of integration,and strong anti-interference capabilities.However,in traditional methods based on millimeter-wave radar point clouds or distance Doppler heat maps,achieving high-precision recognition of small gestures still faces significant challenges.To address this problem,we propose a system called CDCS-CTN(Data Cube Sequence-TD-CNN-Transformer Network),which leverages the richer information in millimeter-wave radar raw data,to extract local features of the data cube sequence by introducing TD(time distribution)wrapper and 3D-CNN(three-dimensional convolution neural network),while retaining the time information of the sequence through a position encoder,and obtaining the global features of the sequence through Transformer network,thereby achieving higher gesture recognition accuracy.Experimental results show that the CDCS-CTN system can achieve a gesture recognition accuracy of 99.75%,significantly higher than other traditional methods.This indicates that our method can better address the challenges of recognizing small gestures by extracting both local and global features of the data,with good practicality and application prospects.Our research focuses on millimeter-wave radar-based gesture recognition:(1)We established a millimeter-wave radar data collection platform and simultaneously captured three different types of gesture data,including LVDS radar echoes,point clouds,and distance Doppler heat maps,using two threads.These data were used to build our dataset,which includes 20 types of micro gestures from 21 participants.(2)We proposed a unique data processing method that converts radar echoes into radar data cube sequences.On the one hand,when using the constant false alarm rate(CFAR)detection method to filter data and obtain point clouds,some useful information may be lost.On the other hand,the Doppler spectrum of the radar lacks information about the target’s angle.Therefore,the received radar echoes are processed into a sequence of data cubes with distance,velocity,angle,and time information to improve the accuracy of gesture recognition.This data processing method fully utilizes the three-dimensional spatial information and time information of gesture movements.(3)We proposed a network based on attention mechanisms,called CTN(TD-CNN Transformer Network),which combines 3D CNN with Transformer in a temporal distribution manner to achieve high gesture recognition accuracy.In this network,multiple 3D CNN synchronously extract features of the data cube,encode them based on their time position in each action,and then feed them into the Transformer layer to obtain long-distance features in gesture data.Our experimental results show that when using data cube sequences as input features,CTN has higher recognition accuracy than CNN and CNN-LSTM,with an accuracy of approximately 99.75%.Therefore,our research provides an efficient,accurate,and reliable solution for millimeter-wave radar-based gesture recognition.