基于规则与机器学习融合的换道决策建模方法研究

Lane-changing Decision Model Development by Combining Rules Abstract and Machine Learning Technique

自主换道系统是智能车辆技术的重要发展方向,而换道决策是自主换道系统的关键。针对结构化道路下的自由换道场景,提出基于规则与机器学习融合的换道决策模型。针对换道决策过程中存在的多参数与非线性问题,提出基于支持向量机的换道决策模型,并引入贝叶斯优化算法确定决策模型的最优参数。从基于规则的角度出发,分析换道决策过程中的影响因素,包括换道必要性、安全性与换道收益,并将上述规则转化为模型训练数据的新特征与安全约束,对原有训练样本进行增广,以提升支持向量机模型的分类准确率。在NGSIM数据集上进行测试验证,结果表明,仅利用周围车辆基本行驶信息进行模型训练,对换道行为预测的准确率为73.05%,而引入基于换道规则计算得到的新特征后,模型预测准确率提升至83.83%。

The autonomous lane-changing system is one of the important research directions of vehicle intelligence at present and lane-changing decision plays a decisive role for the success of lane-changing manoeuvre. An enabling lane-changing decision model is proposed by combing rule abstraction and machine learning model for lane-changing on structured roads. Considering the multi-parameter and nonlinear properties of the lane-changing decision-making process, a lane-changing decision model is first developed based on the support vector machine, and the Bayesian optimization algorithm is introduced to determine its optimal parameters. Then the major influencing factors of the lane-changing decision-making are analysed including the necessity, safety and benefit. These factors are added as the new features into the training datasets so that each original training sample is augmented to improve the training effect. Finally, the prediction accuracy of the developed model is examined based on the NGSIM data. The results show that the prediction accuracies of the trained lane-changing decision models based on the augmented and non-augmented training datasets are 84.22% and 77.19%, respectively. This verifies the effectiveness of the proposed lane-changing decision model and its augmented training method.