Multi-person Gesture Recognition Algorithm Using Millimeter Wave Radar

Millimeter-wave radar gesture recognition is an important means of human-computer interaction,with many advantages such as unaffected by lighting conditions,strong dynamic capture capability,and strong privacy protection.It has received widespread attention in recent years.Multi-person gesture recognition based on millimeter-wave radar aims to recognize multiple gestures that coexist within the radar detection range,which can broaden the application scenarios of gesture recognition,improve its practicality and interaction efficiency in many fields such as intelligent cabins and smart homes,and has important application value and research significance.Gesture number estimation,multi-gesture signal separation,and recognition algorithms are the key elements for implementing multi-person gesture recognition.In this thesis,theoretical research and experimental verification work are conducted around the issues of clutter suppression and feature extraction,multi-gesture number estimation and separation,and multi-feature fusion gesture recognition algorithms involved in multiperson gesture recognition,as follows:1.Researched gesture recognition signal preprocessing and multi-dimensional feature extraction methods.The clutter signal was suppressed using an exponential weighted moving target indicator filter.Based on the 60GHz frequency-modulated continuouswave(FMCW)radar,various distance,speed,and angle feature maps of multiple gestures were extracted using time-frequency analysis and spatial spectrum estimation methods.2.For the problem of mutual coupling of multi-person gesture signals under unknown number of gestures,a multi-person gesture separation algorithm based on independent component analysis(ICA)is proposed.The gesture number is estimated by analyzing the circular distribution of eigenvalues of the covariance matrix of the gesture echo signals based on the Geiringer’s theorem.Then,the estimation of the array signal mixing matrix is achieved by j ointly approximating the diagonalization of the whitened observation signals,which leads to the separation of multi-person gesture signals.3.For the problem of residual feature and difficult efficient fusion of multi-dimensional features after multi-person gesture signal separation,a multi-feature fusion gesture recognition algorithm combining mixed attention mechanism is proposed.The mixed attention module is designed based on channel,spatial,and self-attention mechanisms,which achieves feature fusion,feature enhancement,and feature selection,and realizes accurate recognition of gestures.4.In this thesis,we propose a multi-person gesture recognition network training method based on transfer learning.The method utilizes a large-scale dataset of singleperson gestures to perform network pre-training,and then applies transfer learning using a small-scale dataset of multi-person gestures to achieve a significant improvement in multi-person gesture recognition accuracy.All the proposed methods were verified using FMCW radar measured signals,and experiments showed that the methods proposed in this thesis achieved accurate recognition of multi-person gestures.