Abstract: When police carry out traffic law enforcement using multi-target speed measuring radar, one of the most essential things is to judge which lane each vehicle belongs to, and only in this way the captured pictures can serve as the law enforcement evidence. To achieve lane division purpose, traditional way is to obtain a fixed threshold by manual measurement and sometimes the method of coordinate system rotation is also needed, but this method has a large error with difficulty in operating. A new lane division algorithm called Kernel Clustering algorithm based on Statistical and Density Features (K-CSDF) was proposed, which includes two steps: firstly, a feature extraction method based on statistical feature and density feature was used to process the vehicle data captured by radar; secondly, a dynamic clustering algorithm based on kernel and similarity was introduced to cluster the processed data. Simulations with Gaussian Mixture Model (GMM) algorithm and Self-Organizing Maps (SOM) algorithm were conducted. Simulation results show that the proposed algorithm and SOM algorithm can achieve a lane accuracy of more than 90% when only 100 sample points are used, while GMM algorithm cannot detect the lane center line. In terms of running time, when 1000 sample points are taken, the proposed algorithm and GMM algorithm spend less than one second, and the real-time performance can be guaranteed, while SOM algorithm takes about 2.5 seconds. The robustness of the proposed algorithm is better than GMM algorithm and SOM algorithm when sample points have a non-uniform distribution. When different amounts of sample points are used for clustering, the proposed algorithm can achieve an average lane division accuracy of more than 95%.

Key words: multi-target radar, lane division, statistical feature, dynamic radius, kernel, dynamic clustering

摘要： 利用多目标交通测速雷达进行交通执法时，只有正确地判断出车辆所在的车道，抓拍照片才能作为交通执法的依据。传统的分车道方法主要通过人工测量的固定阈值以及坐标系旋转的方法来达到车道划分的目的，但这种方法误差较大并且不易于操作。基于统计和密度特征的核聚类算法（K-CSDF）分两步进行：首先对雷达获取的车辆数据进行特征提取，包括基于统计特征的阈值处理和基于密度特征的动态半径提取；然后引入基于核的相似性的动态聚类算法对筛选出的有效点进行聚类。通过和高斯混合模型（GMM）算法以及自组织映射神经网络（SOM）算法进行仿真对比表明：当只取100个有效点进行聚类时，K-CSDF和SOM算法能达到90%以上的分车道正确率，而GMM算法不能给出车道中心线；在算法用时上，当取1000个有效点时，K-CSDF和GMM算法用时均小于1s，可以保证实时性，而SOM算法则需要2.5s左右；在算法鲁棒性上，K-CSDF对不均匀样本的适应性优于这两种算法。当取不同数量的有效点进行聚类时，K-CSDF可以达到95%以上的平均分车道正确率。

关键词: 多目标雷达, 车道划分, 统计特征, 动态半径, 核, 动态聚类