Vehicle Target Recognition Based On Radar Signal And Deep Learning

The high range resolution profile reflects the superposition of the target scattering points on the radar line of sight,and contains a large amount of fine structure information of the target.Because of its advantage of easy acquisition and storage,small amount of data,little dependence on observation scene and target movement,it has unique advantages in real-time processing and has received continuous attention in the field of radar target recognition.Therefore,how to obtain robust and reliable features from HRRP for target recognition has always been the key research content in this field.For high range resolution profile signals of the vehicle target,thesis carries out relevant research from two aspects of feature extraction and time-frequency feature fusion recognition:In order to solve the problem that Convolutional Neural Networks（CNN）recognize HRRP with slow convergence,thesis proposes a feature fusion network（TMCNN）which uses the model parameter features to assist CNN network training.The input of CNN network is HRRP,and the characteristics of model parameters are characteristics of target scattering center.According to the geometrical diffraction theory（GTD）,the target electromagnetic scattering model is established,and the parameters of scattering model are solved by using modern spectral estimation method.Then,a parameter feature selection algorithm is designed based on CNN network.By evaluating the parameter contribution of four kinds of parameters,the parameter feature who are most conducive to auxiliary classification is selected splicing into CNN network and fused with time domain feature for joint recognition.In thesis,we compare the TMCNN with some classical deep learning recognition algorithms such as support vector machine（SVM）,stacked auto-encoder（SAE）and long short-term memory（LSTM）.Experimental results show that the model parameter feature splicing and feature fusion algorithm can effectively improve the target recognition rate under different SNR conditions.In order to make the network have stronger feature extraction ability and antitranslation ability,a hybrid model of CNN and bi-LSTM is proposed in thesis.In this model,CNN is also used as a feature extraction method,and Bi-LSTM is embedded in the back end of CNN to replace the original fully connected network for time sequence construction of the proposed features.For further improve the anti-translation ability of the network,the attention mechanism is added to the hybrid model to capture the key regions of feature.Simulation verification shows that proposed model has a great improvement in recognition rate and anti-translation ability.After replacing CNN with TMCNN,the recognition effect of the network is further improved.Compared with the series fusion of the original time-frequency features,the running time of the model parameter feature splicing and fusion algorithm is greatly reduced.