Research On Traffic Target Clustering And Association Method Based On Millimeter Wave Radar Data Sequence Feature

Millimeter wave radar is the main sensor in the field of traffic flow detection,and multi-target track extraction is a hot issue in this field.Unsupervised data clustering algorithm is one of the main technical means to solve this problem.Aiming at the clustering error caused by radar multipath noise and sparse point cloud,automatic target number estimation and target association,the main research work of this paper includes:(1)According to the temporal and spatial variation characteristics of the actual data of traffic targets collected by millimeter-wave radar,this paper analyzes the non-sequential features and the sequential features respectively.In the aspect of non-sequence feature analysis,the maximum likelihood statistical analysis is adopted,and the basis of effective data screening is given.In the aspect of sequence feature analysis,the time series scatter plot is used to analyze the characteristics of sequence features changing with time or space,which lays a foundation for the application of sequence features in the follow-up algorithm research.(2)Based on the actual data phenomenon,this paper divides the environmental clutter into two categories: non-sequential clutter and sequential clutter.Original DBSCAN(Density-Based Spatial Clustering of Applications with Noise)clustering algorithm is limited by its principle and can't deal with sequential clutter.To solve this problem,this paper proposes a DB-CCE(DBSCAN-CCE)algorithm framework.The proposed algorithm combines the Connected Center Evolution(CCE)algorithm with DBSCAN algorithm,adds the frame number feature to distinguish the target trace sequence from the sequential multipath clutter trace,estimates the target number,and adaptively determines the clustering result through the decision graph.The algorithm is verified and analyzed by the data of three real scenes,and the results show that DB-CCE clustering algorithm can achieve high clustering accuracy in various scenes.(3)Aiming at the low real-time problem caused by the high-dimensional association hypothesis space of GLMB(Generalized Labeled Multi-Bernoulli Filter)multi-target tracking algorithm,this paper proposes a multi-sequence feature weighted nearest neighbor(MSFW-NN)association filter algorithm.The DB-CCE algorithm greatly reduces the dimension of the association hypothesis space,and then calculates the association similarity by weighting the sequence features,so that the original nearest neighbor target association algorithm can make use of the information gain brought by the sequence features.Three experiments show that MSFW-NN algorithm has higher real-time performance and similar accuracy than GLMB filter in the scene with moderate target density.