KDM-OPTICS based vehicle clustering algorithm for traffic millimeter-wave radar

Abstract Traffic millimeter-wave radar suffers from environmental noise, dynamic changes in the number of vehicles, and feature similarity leading to degradation of point-trace clustering accuracy when measuring moving vehicle targets. We propose a new density-based clustering algorithm to address this problem. Unlike traditional datasets, the radar measurement dataset used in this algorithm contains specific multidimensional information against a moving vehicle target, such as target distance, normal angle, relative velocity, and reflection intensity. In order to eliminate the correlation interference between the parameters as well as the influence of measurement units, the normalized Mahalanobis distance method is introduced to generate the target spot trace distance matrix. Subsequently, based on the target point trace distance matrix, the reachable distance map is quickly generated by KD tree nearest neighbor search. At the same time, the dynamic adjustment of neighborhood radius of OPTICS is introduced to enhance the contrast of the target point traces in the reachable distance matrix, to obtain the point trace segmentation data of similar vehicles,and to realize the point trace clustering of vehicle targets in the distance and speed dimensions. Finally,the target ranking recognition clustering structure algorithm (KDM-OPTICS) based on KD tree and normalized martens distance is implemented.In this study,experimental validation based on vehicle target data from real urban traffic scenarios is carried out.The results show that enhancing the target information by local power-law transformation can improve the stability and reliability of distinguishing vehicles with similar distance speed. In addition,the KDM-OPTICS algorithm has higher recognition accuracy and lower computational speed compared with other algorithms.