Research On Gesture Segmentation And Recongnition Algorithm Based On Millimeter Wave Radar

Gesture recognition is an important way of human-computer interaction.The gesture recognition system based on millimeter wave radar has the outstanding advantages of small size,low cost,anti-weak light,anti-interference and convenient privacy protection.Radar-based gesture recognition can be divided into isolated gesture recognition and continuous gesture recognition.At present,most researches are on isolated gesture recognition.Isolated gesture recognition requires manual segmentation and alignment in advance,but in practical applications,manual segmentation and alignment cannot meet the needs of actual applications.Therefore,based on millimeter-wave radar,combined with radar signal processing technology and deep learning classification and recognition technology,the continuous gesture segmentation and recognition method is studied.First,the traditional characterization gestures were adjusted,using both fingertip motion information,palm and arm motion information to characterize gestures,and completed the corresponding data collection experiments,as well as spectrum analysis,clutter interference suppression,data Formatting and other preprocessing work.Secondly,according to the characteristics of radar data,a segmentation method of corresponding gesture actions is proposed according to the characteristics of radar data.That is,the sliding window algorithm is used to pre-segment the data stream,cut into fixed frame data units,and then input the data units into the model to predict.According to the probability value of each label predicted by the model,the gesture action area is divided to obtain a relatively accurate time-domain segmentation of the gesture.Finally,the method based on convolutional neural network and recurrent neural network model is combined to build a hybrid network model,that is,first use the three-dimensional convolutional neural network to learn the short-term spectral characteristics of gestures,and then use the recurrent neural network to learn the long-term sequence of gesture data information.Under the experimental conditions in this paper,the recognition accuracy of the hybrid network reaches 97.5%,which is higher than the traditional recognition method using convolutional networks or recurrent networks,and the performance is more stable,which meets the requirements of continuous gesture recognition systems for recognition accuracy.