Research On Clustering Algorithm Of Vehicular Millimeter Wave Radar

Nowadays,millimeter wave radar is widely used in unmanned driving system.With the development of millimeter wave radar technology,the resolution of vehicular millimeter wave radar is getting higher,leading to the increase of the amount of data obtained from the same target and the amount of data in data set.In this way,the state of target can be analyzed more accurately by the unmanned driving system,but the increase of the amount of the data in data set will affect the unmanned driving system to judge the number of targets.Thus,appropriate clustering algorithms to process the target data set to get the number of targets are needed.To vehicular millimeter wave radar,the data density of target data set is uneven and it is difficult to forecast the optimal clustering result in advance because of the complex road environment.There are some deviations between the results of traditional clustering algorithms and the detection scenes.To solve above problems,this thesis studies the clustering algorithms of vehicular millimeter wave radar,which includes the following parts:(1)Summarize traditional clustering algorithms and their problems in the application of vehicular millimeter wave radar.The shortcomings of K-means clustering algorithm,DBSCAN clustering algorithm and Chameleon clustering algorithm in the application of vehicular millimeter wave radar are analyzed by the derivation of algorithm and simulation experiments.The parameter of K-means clustering algorihm is difficult to determine in advance in the complex environment,while DBSCAN clustering algorithm and Chameleon clustering algorithm have poor performance when the data density of target data set is uneven.(2)An algorithm called Vector-Similarity clustering algorithm is proposed.The purpose of this clustering algorithm is to distinguish different targets with close range,velocity and angle by clustering target data.Vector-Similarity clustering algorithm adds the amplitude of spectrum of the IF signal in target data as a new feature information based on the distance,speed and angle feature information.Then,the algorithm processes the target data set by cosine similarity method and Jaccard similarity coefficient method.In the experiment,target data set is from the scene and is clustered by Vector-Similarity clustering algorithm.The performance of the algorithm is verified by analyzing the clustering results and calculating the internal clustering quality indexes.The algorithm is compared with K-means clustering algorithm,DBSCAN clustering algorithm and Chameleon clustering algorithm.The results of experiments and the comparison of clustering quality indexes show that Vector-Similarity clustering algorithm is effective.(3)An algorithm called Ellipse-DBSCAN clustering algorithm is proposed.The purpose of this clustering algorithm is to solve the problem of DBSCAN clustering algorithm in vehicular millimeter-wave radar.Ellipse-DBSCAN clustering algorithm turns the circular neighborhood of DBSCAN clustering algorithm into elliptic neighborhood which is able to change adaptively.Firstly,the algorithm adaptively caculates the short axis parameter and the long axis parameter of the elliptic neighborhood by the feature information of target data and the angle resolution of millimeter wave radar.Then the algorithm analyzes the parameter of density threshold through the characteristics of target data set.Then,according to the flow of DBSCAN algorithm,the algorithm combines with above three parameters to process target data set.In the experiment,target data set is from the scene and is clustered by Ellipse-DBSCAN clustering algorithm.The performance of the algorithm is verified by analyzing the clustering results and calculating the internal clustering quality indexes.The algorithm is compared with K-means clustering algorithm,DBSCAN clustering algorithm and Chameleon clustering algorithm.The results of experiments and the comparison of clustering quality indexes show that Ellipse-DBSCAN clustering algorithm is effective.