Ground Slow Moving Target Classification And Refined Recognition Method Based On Deep Learning

Ground slow moving target recognition is an important part of radar target recognition,and has great application value in the field of military surveillance and security defense.Pedestrians and vehicles are the main monitoring objects of ground surveillance radars,of which a variety of micro-movements can provide effective information for the discrimination of types and motion status of ground slow moving targets.Therefore,it is one of the current research hot topics to utilize the micro-Doppler characteristics to recognize the ground slow moving targets.At the same time,with the development of algorithms and the improvement in the efficiency of graphics computing units,deep learning has been gradually applied in the field of radar target recognition.Therefore,it is of great significance to study the classification and refine recognition of ground slow moving targets based on deep learning.The sample features can be automatically extracted by the deep learning method,but there is a lack of physical support for the features.Therefore,how to design a fusion network effectively combining the physical features and the deep network to improve the recognition accuracy of ground slow moving targets is an urgent problem.It is difficult to obtain the mearsured data of ground slow moving targets because of the difficulty and the long cycle of admission,and easy to encounter over-fitting problems in the training process.It is necessary to augment the existing datasets to make the deep network fully learn the deep level features of the sample and improve the recognition accuracy.In addition,it is important to finely identify their micro-moving parts,so as to provide strong support for the extraction of motion parameters and the prediction of motion state.This dissertation is funded by the Military Science and Technology Commission project and the Youth Program of National Natural Science Foundation of China.With the practical application requirements of target detection and recognition for ground surveillance radar systems,ground slow moving target classification and refined recognition based on deep learning,is carried out.The research focuses on the ground slow moving target recognition based on the deep fusion network,radar echo data enhancement based on Generative Adversarial Network(GAN)and the refined recognition for pedestrian micro-motion parts based on improved Fully Convolutional Network(FCN).The main research contents include:1.Pedestrians and vehicle targets have rich micro-movements.Different micro-movements correspond to different micro-Doppler characteristics.Therefore,it is an important premise to study the classification and recognition of ground slow moving targets by utlizing the difference of micro-Doppler characteristics.Firstly,a pedestrian radar echo modeling method based on improved biomechanical model and a pedestrian gait radar echo generation method based on motion capture data and electromagnetic scattering characteristics are proposed.Secondly,time-frequency analysis is utilized to analyze the micro-Doppler characteristics of them;Finally,three dataset of the time-frequency spectrograms of the ground slow moving targets are establised,including pedestrian armed or not database based on improved biomechanical models,pedestrian gaits database based on motion capture data and electromagnetic scattering characteristics,and measured pedestrian and vehicle database based on continuous wave radar.The construction of the time-frequency spectrogram dataset for the ground slow moving target has laid the foundation for the subsequent research on the ground slow moving target recognition based on deep learning,the ground slow moving target echo enhancement algorithm,and the refined identification of pedestrian micro-movement parts.2.Aiming at the problems that the features extracted by deep learning algorithms are not interpretable and the effectiveness is not guaranteed,a ground slow moving target recognition algorithm based on deep fusion network is proposed.Firstly,manual feature extraction is performed on radar echo samples to form a one-dimensional feature vector;Secondly,a convolutional neural network(CNN)is used to construct a target recognition network;Then,a feature fusion structure is introduced into the network,the features extracted manually and the features extracted by deep learning are fused through a fully connected layer;Finally,the performance of the recognition algorithm is evaluated using the test set.The effectiveness of the proposed algorithm is verified by comparing the experimental results of the original network and the proposed deep fusion network under different training set ratios and different signal-to-noise ratios.3.Aiming at the difficulty of data acquisition and the difficulty for existing slow-speed target sample database to meet the requirements of deep learning which result in the lack of generalization ability of the model,a data enhancement algorithm for ground slow moving targets based on Wasserstein generation adversarial network(WGAN)is proposed.Firstly,the basic principles of GAN are studied,and the reasons for the instability of GAN training are analyzed from the perspective of loss function;Secondly,a radar data enhancement network for ground slow moving targets based on WGAN is built,the original dataset was expanded based on the generated mode;After that,the distribution difference between the generated sample and the real sample were measured by image evaluation indicators;Finally,the target recognition performances based on the original sample and the enhanced sample were compared under different signal-to-noise ratios.The experimental results show that the generated samples have learned the data distribution of the real samples to some extent,and the effectiveness and robustness of the recognition algorithm are promoted by using the original dataset which was enhance by the generated samples.4.Aiming at the challenge that pedestrian micro-Doppler signals overlap with each other in the time-frequency domain and the difficulty of the refined recognition,a refined recognition algorithm for pedestrian micro-motion parts based on improved FCN is proposed.Firstly,on the basis of the original pedestrian asynchronous database based on biomechanical models and the pedestrian asynchronous database based on motion capture data and electromagnetic scattering characteristics,the micro-Doppler time-frequency dataset of 12 micro-moving parts of pedestrians is established which serves as a labeled data set for fine identification;Secondly,a refined recognition network for pedestrian micro-motion parts based on improved FCN is constructed to separate and refine the micro-Doppler signals of each micro-motion part of pedestrians.Finally,in order to evaluate the effectiveness of the proposed refined recognition algorithm quantitatively,several evaluation indicators are used to calculate the difference between the predicted and real separated signals.