ABSTRACT

A lane-changing process is complicated due to multiple factors in the driving environment, and unsafe lane-changing behaviour may lead to a severe crash. This study proposes a method for the driving angle prediction of lane changes based on extremely randomized decision trees. First, the harmonic potential is defined to characterize the interaction between the lane-changing vehicle and the surrounding vehicles. Next, we construct extremely randomized decision trees to predict driving angles considering relative velocity, relative acceleration, and potential as input variables. Then, the NGSIM dataset is used to verify the method proposed, and the lane-changing process is divided into two stages by different environments. Furthermore, a comparison of prediction performance with several traditional machine learning methods further demonstrates the superior learning ability of the proposed method. Finally, we conduct a sensitivity analysis on the significant variables and discuss the effects of these variables on the prediction results.

摘要

行车环境中的多种因素导致换道过程复杂，不安全的换道行为可能导致严重的碰撞事故。本研究提出了一种基于极度随机化决策树的变道行驶角度预测方法。首先，定义谐波势来表征换道车辆与周围车辆之间的相互作用。接下来，我们构建了极其随机的决策树来预测驾驶角度，将相对速度、相对加速度和势能作为输入变量。然后，使用NGSIM数据集对提出的方法进行验证，根据不同的环境将变道过程分为两个阶段。此外，与几种传统机器学习方法的预测性能比较进一步证明了所提出方法的优越学习能力。最后，我们对显着变量进行敏感性分析，讨论这些变量对预测结果的影响。