基于KL散度的驾驶员驾驶习性非监督聚类

Unsupervised Clustering of Driving Styles Based on KL Divergence

为深入理解不同驾驶员的驾驶行为特点,本文中提出了一种基于KL散度的驾驶员驾驶习性非监督聚类算法。首先,建立了驾驶员驾驶数据实车道路试验采集平台,对84位驾驶员进行了测试;接着,将每名驾驶员的驾驶数据视为一个高斯混合模型(GMM),采取EM算法对其进行参数估计;最后,通过蒙特卡洛算法对各GMM之间的KL散度进行估计,从而获得不同驾驶员差异性的定量描述,将驾驶员聚为不同习性类别。对聚类后各类驾驶员的驾驶数据的统计分析表明,所提出的非监督聚类算法能有效实现不同驾驶习性驾驶员的聚类。

In order to understand the driving style features of different drivers, an unsupervised clustering algorithm for the driving styles of drivers is proposed in this paper based on Kullbaek-Leibler （KL） divergence. Firstly an acquisition platform for the driving data of drivers in real vehicle test is built, and 84 drivers are tested. Then the driving data of each driver are regarded as a specific Gaussian mixture model （GMM） , and whose parameters are estimated by using expectation maximization algorithm. Finally Monte Carlo algorithm is employed to esti- mate the KL divergence between GMMs, hence the quantitative description on the discrepancies of different drivers is obtained and drivers are clustered into different eatagories of style. The results of statistical analysis on the driving data of drivers in each eategory alter clustering show that the unsupervised clustering algorithm proposed can effectively achieve the elustereing of drivers with different driving styles.