Research On The Vehicular Radar Tracking Association Method Based On Machine Learning

With the rapid development in recent decades,Chinese car ownership has increased rapidly.How to ensure driving safety in a complex traffic environment is a problem that deserves attention.At the same time,the vehicle-mounted millimeter-wave radar has a complex working environment and dense targets.In order to solve the problem of the performance degradation of traditional radar track association algorithms under the condition of heavy clutter,this paper proposes track association algorithms,which uses the information mining and learning capabilities of machine learning.Firstly,the commonly used radar track initiation algorithm and track maintenance algorithm are studied,and the performance of these algorithms is analyzed and compared through simulation experiments.Based on the classification capability of machine learning algorithms under the framework of track quality management,an vehicular radar track initiation method based on machine learning is proposed.The basic idea is to extract valid information from massive experimental data for classification of radar measurements,and reduce the effect of false measurements on the performance of the track initiation algorithm.Then,the preprocessing of radar experimental data sets and the training process of classifiers are explained in detail.This thesis select several typical machine learning classifiers,including KNN,decision tree,SVM and neural network,study their principles and conduct experimental comparisons in terms of model training time,single measurement classification accuracy,algorithm application performance,etc.,and analyze the their adaptability in radar track initiation algorithm.Finally,in view of the shortcomings of the traditional track maintenance algorithm,the weighted sum of the multi-frame measurement prediction states is used as the cluster center of the fuzzy clustering algorithm,the membership between the cluster center and candidate measurements are used as a basis to reduce the number of associated hypotheses.In this way,the real-time performance and the stability in complex situations can be improved.Because the multi-frame measurement information is extracted,the incorrect tracking maintenance in the case of multiple objects crossover due to a single wrong connection is reduced.The simulation experiment results show that the track association algorithm based on machine learning proposed in this paper is more robust and stable in complex environments.