UWB Radar Human Motion Recognition System Based On PointNet

Ultra Wideband(UWB)radar is widely used in the field of signal acquisition and processing because of its strong anti-interference and high range resolution.In the field of human motion recognition,the recognition rate of traditional micro-Doppler effect is greatly affected by the image resolution of time-frequency image,and the aliasing of the motion feature information of different targets in the multi-person scene will affect the recognition effect.In addition,the time-frequency image only contains the velocity information of the moving target extracted from the radar signal at different times,and the distance information is omitted,resulting in information waste.Therefore,simultaneously extracting the time,distance and velocity information of the target and making full use of it can help improve the effect of human motion recognition.In this thesis,echo signals are collected by modeling human body and simulating the process of radar detection of moving targets.The motion features of human body are expressed as point clouds by extracting time,range and velocity information.In the multi-person scene,the motion feature data of different targets only have little overlap in the three-dimensional space,which can be easily separated and then recognized.Because deep learning has shown strong ability and potential in the field of pattern recognition,this thesis selects Point Net model suitable for point cloud data processing to recognize the motion types of different human targets.The specific research contents are summarized as follows:Firstly,the human body is modeled based on the human skeleton data in the CMU MOCAP dataset,and then the echo signal of UWB pulse radar is simulated to detect human movements.Then simulate the process of UWB pulse radar detecting human action and collecting echo signal.Through pulse compression technology and Range-Doppler processing,the distance and velocity information of motion at different times are obtained from the echo signal.The proposed Twice mean filtering method can quickly and high-quality extract the effective target from the Range-Doppler image.After arranging the extracted data according to the sequence of time frames,the point cloud data expressing the features of human motion can be obtained.Secondly,since human motion detection in daily life is not always conducted for a single person,it is very important to separate the data representing different target motion fratures in the same point cloud for subsequent motion recognition processing.Since the number of targets is variable,a mean clustering method which can automatically determine the number of clusters is proposed by combining the density based spatial clustering of applications with noise(DBSCAN)algorithm and K-means algorithm.The DBSCAN algorithm is used to analyze the data of each single frame to determine the number of target people.When the number of target is greater than 1,the K-means clustering algorithm is used to separate the motion features of multiple targets.For the clustering algorithm based on distance mean is not good at dealing with the problem of non-convex point cloud,the complete point cloud data can be divided into multiple subspaces along the time dimension.Then the mean clustering is carried out respectively and the clusters belonging to the same target are grouped into a large cluster,so as to obtain the result of different target data separation.Finally,using the point cloud data obtained from the two processes of feature extraction and target separation,the motion feature data set is established,including five actions:walking,forward jumping,jogging,climbing ladders and climbing stairs.Each motion has1200 data samples,which are divided into training set,verification set and test set according to 4:1:1.Build pointnet network,use training set and verification set to train and optimize the network model,and test the model performance on the test set.The results show that the human motion recognition system based on Point Net model and UWB pulse radar proposed in this thesis can achieve good recognition effect in single or multi-person scenes.