Research On Classification Method Of Human And Vehicle Targets Based On Millimeter Wave Radar

Because of its high range resolution and speed resolution,millimeter wave radar has gradually entered people's field of vision in recent years and has begun to play a role in fields including autonomous driving and smart home.The existing target classification methods based on millimeter-wave radar obtain the target's time-frequency map,Range Doppler(RD)map,slice-level point cloud or other types of data from the target echo after preprocessing,and then extract the features that can reflect the target's micro-motion,macro-distance or other information from different feature domains and input the feature vector into different classifier to complete the target classification.However,under normal circumstances,the target echo can only reflect the target's current state and characteristics,the reflection to temporal sequence information is not obvious.Relatively,the combined use of longer target echoes can make fuller use of target detail information and temporal sequence information to a certain extent,more comprehensively reflect the state and characteristics of the target's entire movement process,and can increase the accurateness and robustness of classification.In this paper,the research on the application of millimeter-wave radar echo in the classification of people and vehicles is carried out,and has carried out the following work:1.The third chapter of the thesis perform simulation analysis on the millimeter wave radar echo of the typical targets,including: first build different motion models for three types of targets: pedestrian,four-wheeled vehicle,and two-wheeled vehicle,and simulate the target echoes under these models according to the basic principle of millimeter wave radar.After that,using the obtained millimeter-wave radar simulation data,the target characteristics are analyzed through time-frequency maps and RD maps,showing that it is feasible to use the micro-motion and macro-range characteristics of the millimeter-wave radar to classify the above three types of targets.Finally,the factors affecting the target radar echoes and the impact of these factors on the classification task are analyzed.2.The fourth chapter of the thesis proposes a feature extraction method based on the target RD map.The method extract features that reflect the target's micro-motion and macro information from a single frame of RD map,and extract features that reflect target temporal sequence information from two frames of RD maps,and finally the final feature vector is obtained by splicing.Compared with the traditional single-frame feature extraction method,this method utilizes the temporal sequence information of the target,and can extract richer features and achieve a better classification effect.Through related experiments on measured data,it is proved that using the features proposed by this method to perform experiments has a certain improvement in classification accuracy compared with traditional single-frame feature extraction methods,which verifies the effectiveness of this method in target classification based on millimeter wave radar echoes.3.In order to make full use of the information in the target echo to improve the overall classification accuracy,Chapter 5 of the thesis proposes a target classification method by combining Convolutional Neural Network(CNN)and Long Short-Term Memory(LSTM).Different from Chapter 4,only the first and last two frames of RD maps are used to obtain the temporal sequence information of the target.This method first uses a CNN to obtain the feature vector of each frame through convolution operation on each frame of the RD map,and then use the feature vector of each frame RD map as the input of each time step of LSTM.The state of the last hidden layer is used to obtain the class prediction of the target.CNN can learn rich target detailed information,and avoid the waste of some information in the manual feature extraction method.The use of LSTM can make full use of the detailed information of all frame RD maps,and the use of temporal sequence information will be more fully utilized.Experimental results based on measured data show that the classification accuracy of this method is improved by more than 6% compared with the feature extraction method proposed in Chapter 4,which verifies that the use of CNN for feature extraction and the joint use of multi-frame RD maps to mine and utilize temporal sequence information are effective.