Target Recognition Based On FMCW Millimeter-Wave Radar

Frequency Modulation Continuous Wave(FMCW)millimeter-wave radar has the characteristics of low cost,high detection accuracy,all-weather,all-time working,etc.It is widely used in civil radar field at present,and radar target recognition is an important function of radar.However,at present,the civil FMCW millimeter-wave radar has a low recognition accuracy when identifying road targets due to the complex road environment and excessive interference.There is still a long way to go before radar target recognition technology can be really applied to real roads.Therefore,this paper conducts research on target recognition based on FMCW millimeter-wave radar,and the main work contents are as follows:1.The principle of FMCW millimeter wave radar is analyzed.Firstly,the FMCW millimeter-wave radar system with sawtooth waveform is introduced,and based on this,the Intermediate Frequency(IF)signal of the radar is analyzed.Then,the ranging,velocity and Angle measurement principles of FMCW millimeter-wave radar are analyzed,which lays a theoretical foundation for future work.2.The algorithm of radar target recognition is studied.Radar target recognition is mainly divided into machine learning and deep learning algorithms.Firstly,radar target recognition based on Machine learning is studied.In Machine learning,K-nearest Neighbor(KNN)classification algorithm and Support Vector Machine(SVM)classification algorithm are mainly studied.The experimental results show that in this experiment,when K=3 and Manhattan distance is selected as the range measurement method,the KNN algorithm has the highest average recognition rate of radar target recognition,about 63.44%.Although the average recognition rate is not high,its training and recognition speed is the fastest.In this experiment,the highest average recognition rate of radar target recognition using SVM algorithm is about 62.38%,and the training and recognition speed of this algorithm is the slowest.Then,radar target recognition based on deep learning is studied.In deep learning,the Convolutional Neural Network(CNN)classification algorithm is mainly studied.The experimental results show that the average recognition rate of the CNN algorithm using twolayer convolution layer and pooling layer structure for radar target recognition is about 66.11%,and the average recognition rate of the CNN algorithm using three-layer convolution layer and pooling layer structure for radar target recognition is about 64.26%.The training and recognition speed of CNN algorithm is between KNN algorithm and S VM algorithm.3.An improved CNN classification algorithm is proposed.As the CNN classification algorithm tends to cause the internal covariate shift(ICS)problem,the Batch Normalization(BN)layer is added to the CNN algorithm to suppress the influence of ICS problem and improve the average recognition rate.Proposes three optimization methods,in view of the two kinds of structures of CNN algorithm made six experiments,the experimental results show that,in this experiment,adding a BN layer between each convolution layer and the ReLU activation function of the CNN algorithm with two convolution layers+pooling layers is the best.The average recognition rate of the optimized algorithm is the highest and the training and recognition speed is fast,with an average recognition rate of 71.23%.In terms of the average recognition rate,the proposed optimization algorithm improves by about 9%compared with the machine learning algorithm,but in terms of training and recognition speed,the proposed optimization algorithm is sacrificed compared with the KNN algorithm.The reason for the improvement of the average recognition rate is that,compared with other algorithms,adding BN layer to CNN algorithm makes the data distribution in the training and recognition process more stable,so as to better extract the stable features of the target and improve the average recognition rate.