Parallel Implementation Of Radar Target Recognition Algorithm Based On CPU And GPU

Modern warfare puts forward higher requirements on the level of radar automation and intelligence,and the intelligent information processing technology represented by radar target recognition has received extensive attention.With the continuous improvement of radar signal bandwidth and the increasing number of target types in the recognition database,it has brought great challenges to complete the target recognition task in real time.Because radar target recognition tasks have a good parallel structure,efficient parallel processing algorithms have become a research hotspot in the field of target recognition technology.At the same time,the parallel processing capabilities of the hardware processors represented by the central processing unit(CPU)and the graphics processing unit(GPU)are becoming more and more powerful,which provides the possibility of parallel acceleration of radar target recognition algorithms.Based on the above background,this thesis carried out the parallel design research of radar high resolution range profile(HRRP)recognition-related algorithms,and implemented the algorithm on the multi-core CPU processor hardware platform and CPU+GPU heterogeneous platform.The main work is summarized as follows:1.Combined with the HRRP recognition process of the linear frequency modulation pulse radar system,the principle of the commonly used algorithm in each link is explained.First,the segmented pulse compression algorithm for pulse compression processing of large-time and wide-bandwidth signals is introduced;then the Keystone transform and its two common implementation methods for correcting the target moving distance over unit during coherent accumulation are introduced,namely the DFT+IFFT algorithm and Chirp-Z algorithm;then introduces common pre-processing methods for HRRP data sensitivity;finally introduces four classic statistical recognition models and convolutional neural network recognition models.2.The hardware structure of the CPU and GPU processor is described.Then the C++11 multithreading programming method for parallel programming of multi-core CPU processors and the compute unified device architecture(CUDA)for parallel programming of NVIDIA GPU processors and the programming,execution model of CUDA are given.Finally,the debugging analysis methods and kernel optimization techniques of CUDA parallel programming are given.3.The parallel structure of each algorithm in the HRRP recognition process is analyzed in detail,the corresponding parallel implementation scheme is designed,and the single-thread implementation and multithreading parallel implementation of each algorithm under the CPU platform and multithreading parallel implementation under the CPU+GPU platform are completed.Through simulation experiments,the running results of the algorithms in the two parallel implementations are verified,and the acceleration effect of the two parallel implementations is compared and analyzed based on the CPU single-thread running time.Among them,segmented pulse compression algorithm,DFT+IFFT algorithm,Chirp-Z algorithm,iterative alignment algorithm,maximum correlation coefficient(MCC)classifier and adaptive Gaussian classifier(AGC)modeling,MCC recognition,AGC recognition,factor analysis(FA)modeling and complex factor analysis(CFA)modeling process' s CPU multithreading implementation can achieve 3 to 5 times acceleration effect,the corresponding GPU parallel implementation can also achieve more than 5 times acceleration effect;In addition,based on the recognition method of convolutional neural network has also achieved a certain acceleration effect.This shows that the multithreading parallel implementation based on CPU and the parallel implementation based on CPU+GPU heterogeneous platform can significantly improve the real-time performance of radar target recognition tasks.