Research On Target Recognition Method In Millimeter Wave Radar Scenario Of Internet Of Vehicles Based On Deep Learning

Millimeter-wave radar are widely used in autonomous driving and intelligent transportation systems due to their low cost,high integration,high detection accuracy,and the ability to work around the clock.In the intelligent vehicle road coordination system,the accurate perception and analysis of traffic environment information is the key factor to realize the coordination of people,vehicles and roads,and the target recognition and classification technology based on millimeter-wave radar can provide strong technical support for traffic environment perception.Based on the deep learning method,this paper studies the recognition and classification of targets on the road in the millimeter wave radar scene.The main contents and contributions are as follows:1.On the basis of studying the working principle of millimeter-wave radar and typical target feature information,for the problem of single target recognition,a multi-channel regional Range-Velocity spectrum is constructed as the input feature data of the convolutional neural network.The feature data can be separated the influence of energy scattering of different targets can be suppressed,the errors caused by the fluctuation characteristics of the target radar cross section and the noise of the hardware platform can be suppressed,consequently,the recognition performance can be significantly improved.The simulation results show that the recognition accuracy of the convolutional neural network based on the multi-channel regional range-velocity spectrum for the four types of objects on the road(namely,buses,cars,electric vehicles and pedestrians)is 95.1%,90.2%,93.2% and 92.6%,respectively.In terms of recognition accuracy,convolutional neural networks are far superior to traditional machine learning methods,such as support vector machine and K-nearest neighbor algorithms.2.The basic architecture of convolutional neural networks and related network parameter optimization algorithms are studied.After having an in-depth understanding of the working principle of the convolutional neural network,in order to reduce the computational time and parameter storage of the convolutional neural network model,from the perspective of the convolutional layer,a cosine similarity index is proposed to measure the convolutional neural network.The cosine similarity is used to evaluate the degree of repetition between the kernels,and a method to simplify the network structure based on the cosine similarity between the convolution kernels is subsequently proposed.The simulation results show that this method can reduce the computation and parameter storage consumption of the convolutional neural network by 74.67% and 73.71%,respectively,while ensuring that the recognition performance index of the convolutional neural network is almost not decreased.3.After studying the target recognition principle based on the YOLOv3(You Only Look Once version3)algorithm,in order to apply the YOLOv3 method to recognize targets in the millimeter-wave radar scene,the full RV spectral feature data containing multi-target information in the millimeter-wave radar scene is constructed.Based on the recognition of multiple targets,the clustering of scattering points of different targets can also be realized.The simulation results show that the mean average precision indicators for the detection of three types of objects on the highway,such as cars,light trucks and heavy trucks,can reach 94.6%,97.2% and 95.0%,respectively.