Research On Human Target Recognition Method Of Millimeter Wave FMCW MIMO Radar

Millimeter wave radar has been widely used in intelligent driving,patient monitoring,traffic control and other fields due to its good penetration ability,high detection accuracy.On the other hand,commercial products which are convenient,and compact also make contributions on it.In the existing researches on human recognition which are based on millimeter wave radar,the use of point cloud does not take into characteristics that density differs in range,and most researches are based on the single indirect characteristic,so its characteristics are not fully utilized to complete the classification and recognition tasks.In this way,research on the above problems is carried out by adopting methods based on machine learning and deep learning.We complete the clustering,behavior classification and skeleton posture reconstruction of human,and the specific contents of this paper are as follows:1.Aiming at the problem of different density of point cloud targets in millimeter wave radar,a robust adaptive 3D point cloud clustering method based on density clustering was proposed.Firstly,according to the FMCW MIMO signal processing theory,the distance,azimuth and pitch information of the target were extracted from the original data.Secondly,combined with the range resolution and angle resolution of radar,the extracted threedimensional information was expressed in the form of voxels,and the corresponding local measurements of each voxel were calculated.Thirdly,the clustering search area of each voxel was calculated according to the local measurement.Finally,combined with genetic algorithm,the best parameters needed in the clustering process was adaptively found,and clustering was realized.The results show that the method can cluster 3D point cloud obtained from millimeter wave radar effectively.2.Aiming at the problem that single feature based methods of radar target classification are unable to take consideration of target features fully,a model of multimodal classification is proposed.The model fuses point cloud and micro-doppler information of the target,namely to consider the physical shape,spatial structure characteristics and micro characteristics of the target.Firstly,point clouds and micro doppler spectra of target are obtained from echo signal by different processing.Then image recognition and point convolution networks are used respectively on the doppler spectra and point cloud for feature extraction.Next,bilinear pooling is used to fuse outputs of two channels.Finally,features are inputted to the fully connection layer and classifier to achieve the category of target.The test results of dataset show that compared with the target classification based on single feature,the accuracy of multi-mode model is improved by 3.09%～8.59%.3.Aiming at the problem that current methods of targets reconstruction can reconstruct human skeleton pose,but they lose the spatial information or don’t take density of point cloud into consideration.We propose a skeleton pose reconstruction method that combines point convolution to extract features from point cloud.By extracting the local information and density of each point in point cloud of the target,the spatial location and structure information of the target can be obtained,and the accuracy of the pose estimation is increased.The extraction of point cloud features is based on point-by-point convolution,that is,different weights are applied to different features of each point,which also increases the nonlinear expression ability of the model.Experiments show that the proposed approach is effective.We offer more distinct skeletal joints and a lower mean absolute error,average localization errors of 6.1 cm in X,3.5 cm in Y and 3.3 cm in Z,respectively.