Target And Clutter Clustering In Radar Based On Machine Learning

In recent years,more and more unmanned aerial vehicles(UAVs)are abused for illegal activities.Besides,the frequency of bird strikes is increasing year by year.UAVs and birds are both non-cooperative low-altitude slow small targets(LSSTs).Flying at a low altitude,LSSTs are likely to be in the shade of terrain.Additionally,slow speed and small echo intensity make it hard to detect and track LSSTs.Apart from that,LSSTs are in a complicated realistic environment with diverse clutter and interference,so it is impracticable to judge large amount of data artificially,which increases the difficulty of processing.Therefore,this paper implements clustering algorithms in machine learning to distinguish targets from clutter and realizes intelligent radar processing.This paper mainly focuses on the binary clustering of targets and clutter in the urban area or airport.The prime work is as follows:Firstly,the data set is constructed based on the measured data and the appropriate input features are selected.After analyzing the parameters and signal processing flow of the radar system,we get the information in the output of signal processing.Then,the input features are analyzed and verified on their rationality.The input features were pre-processed by standardization in the end.Secondly,targets and clutter are clustered by K-means and density peaks clustering method.We analyze K-means method and density peaks clustering method from the following three aspects: evaluation index,selection criteria of clustering centers and algorithm procedures.Then we use these two methods to do clustering experiments on the data set.The experimental results are analyzed and compared.Thirdly,considering that the data set is class-imbalanced,a density peaks clustering method based on optimized AdaBoost algorithm is proposed.We under-sample the clutter dots and then combine AdaBoost algorithm based on the asymmetric misclustering cost with density peaks clustering method.The experimental results show that the optimized method can effectively improve the identification of target.Finally,to solve the problem of small number of target,a generation method of target tracks based on Generative Adversarial Network is proposed.Firstly,we construct the generation and discriminant model according to the characteristics of the input target tracks.Then the results of network training are obtained under different super-parameters and the optimal group are selected.Finally,we get the generated track.