摘要 针对机器学习跟驰模型中传统轨迹层面误差分析方法考虑因素不够全面、原始数据存在部分不合理驾驶轨迹且难以剔除的问题,提出一种从总体数据角度评价跟驰轨迹的方法——多元高斯模型,将车头间距与相对速度进行联合考虑更能体现个人的驾驶行为。运用高斯分布描述总体数据的概率分布,计算跟驰轨迹的概率值作为对比依据。选用K近邻模型进行轨迹仿真,在NGSIM数据集上进行训练和仿真测试,仿真实验结果表明,多元高斯模型对轨迹的评价更合理,该方法也可用于改善和扩充数据集,剔除部分不合理轨迹,将仿真概率较高轨迹加入到数据集。 In the machine learning car-following model,the traditional trajectory-level error analysis method considers factors insufficiently,for which there are some unreasonable driving trajectories in the original data set that are difficult to eliminate.The paper takes the headway and the relative speed into consideration jointly,and proposes a method to evaluate the car-following trajectory from the perspective of overall data-the multivariate Gaussian model,which can better reflect individual driving behavior.The Gaussian distribution is used to describe the probability distribution of the overall data,which is used to calculate the probability value of the car-following trajectory as a basis for comparison.The K-nearest neighbor model is selected for trajectory simulation,and then training and simulation tests are performed on the NGSIM data set.The simulation experiment result shows that the multivariate Gaussian model is more reasonable for the evaluation of trajectories.This method can also be used to improve and expand the data set,remove some unreasonable trajectories,and add trajectories with higher simulation probability to the data set.
机构地区 合肥工业大学汽车与交通工程学院
出处 《交通科技与经济》 2021年第3期22-29,共8页 Technology & Economy in Areas of Communications
基金 国家自然科学基金面上项目(52072108)。
关键词 跟驰模型 轨迹评价 机器学习 NGSIM 多元高斯模型 car-following model trajectory evaluation machine learning NGSIM multivariate Gaussian model
