Research On Radar Vehicle Target Recognition Based On Deep Learning

The classification and identification of traffic vehicle targets is the core supporting technology of urban intelligent traffic management system,decision-making automatic driving and integrated traffic information platform.In recent years,with the continuous improvement of radar technology and the emergence of various new technologies,vehicle target recognition technology has gradually changed from image processing technology based on optical sensor to radar detection technology with stronger adaptability to weather and environment and more reliable performance.Nowadays,the classification technology of vehicle target using radar has become a hot spot.The traditional radar target classification recognition technology based on statistical model has some shortcomings such as poor classification reliability and low recognition rate due to the mismatch of the model.The deep learning method can improve the model mismatch,and improve the reliability and accuracy of classification and recognition.In the research process of radar vehicle target recognition based on deep learning,because of the variety and types of vehicles and the complex and changeable traffic environment,the total number of vehicle target data sets is insufficient or the data quantity of each category is different greatly,which affects the accuracy of vehicle identification.Based on the data collected by the radar based on the combination of multi input and multi output orthogonal frequency division multiplexing(MIMO-OFDM),this paper studies the radar vehicle target recognition algorithm based on depth learning,including the following main contents:Firstly,through the data processing and data analysis of the vehicle target collected by MIMO-OFDM radar,the two channel depth residual network(2RC-NET)is designed based on the depth learning theory to realize the effective recognition of radar vehicle target.The 2RC-NET model includes three parts: shallow feature extraction,deep feature extraction and feature fusion.The cross layer and cross-channel residual fusion are adopted to reduce information loss and fully extract vehicle target information.Secondly,aiming at the problem of vehicle target recognition in small sample radar,based on the principle of twin network,SC-Siamese is proposed.Different convolution cores of different sizes are set,and the depth separable convolution is introduced,and less parameters are used to suppress the fitting,so as to improve the recognition rate of radar vehicle targets in small sample cases.The network is verified by comparing with other network models and the experiments of vehicle target recognition in non library.Finally,for the problem that the target sample size of different kinds of vehicles is unbalanced and the recognition accuracy is low,three solutions are proposed,including translation data amplification,weight balance,and data enhancement scheme based on deep convolution generative countermeasure network(DCGAN).The DCGAN solution designed has the best effect,which can effectively improve the recognition rate of vehicle target in the case of data imbalance.This thesis studies vehicle recognition based on depth learning method,designs a new neural network model and studies different sample recognition,which provides a way of thinking for the following radar vehicle target recognition.