• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2020, Vol. 56 ›› Issue (16): 155-165.doi: 10.3901/JME.2020.16.155

• 运载工程 • 上一篇    下一篇

扫码分享

基于真实驾驶数据的运动基元提取与再生成

王博洋, 龚建伟, 张瑞增, 陈慧岩   

  1. 北京理工大学机械与车辆学院 北京 100081
  • 收稿日期:2019-11-06 修回日期:2020-03-03 出版日期:2020-08-20 发布日期:2020-10-19
  • 通讯作者: 龚建伟(通信作者),男,1969年出生,博士,教授,博士研究生导师。主要研究方向为无人驾驶车辆、机器学习、运动规划与控制、车辆动力学建模和环境感知。E-mail:gongjianwei@bit.edu.cn
  • 作者简介:王博洋,男,1991年出生,博士研究生。主要研究方向为无人车运动规划和控制。E-mail:wbythink@163.com
  • 基金资助:
    国家自然科学基金资助项目(61703041)。

Motion Primitives Extraction and Regeneration Based on Real Driving Data

WANG Boyang, GONG Jianwei, ZHANG Ruizeng, CHEN Huiyan   

  1. School of Mechanical Engineering, Beijing Institute of Technology, Beijing 100081
  • Received:2019-11-06 Revised:2020-03-03 Online:2020-08-20 Published:2020-10-19

摘要: 类人驾驶系统是通过学习人类驾驶员知识与经验来提升无人驾驶系统适用性与接受度的重要技术途径。为解决驾驶员轨迹和操控层面经验的表述问题,以采集得到的大量真实驾驶数据为依托,提出一种基于轨迹基元与操控基元的分层式驾驶员经验表述模型。轨迹基元以动态运动基元算法进行表征,并由概率提取算法完成基元从无标签连续轨迹数据中的分割提取。操控基元在轨迹基元的提取分类结果上,利用高斯混合模型完成基元的训练,并利用高斯回归算法完成转向操控序列的预测。结果表明,概率提取算法既利用到了表征与提取之间的相互关联关系,又借助于初始分割点的合理设置,提升了算法的效率并使得提取得到的运动基元符合特定的驾驶假设。此外,所提出的运动基元既能以较高精度完成对驾驶员轨迹和操控层面数据的表征,又具备良好的泛化能力以应对运动基元再生成时在期望位置和时间尺度上的变化需求。最终构建了描述全工况驾驶行为的运动基元库,并大幅提升了运动基元应对不同行车环境时的适用性。

关键词: 智能车辆, 运动基元, 驾驶数据, 动态运动基元(DMP), 高斯混合模型(GMM), 高斯回归算法(GMR)

Abstract: The human-like driving system is an essential technical way to improve the applicability and acceptance of an unmanned driving system by learning the knowledge and experience of human drivers. In order to solve the driving skills representation problem at trajectory and control level, by utilizing a large amount of collected real driving data, a hierarchical driver model based on trajectory primitives and operation primitives is proposed. The trajectory primitives are represented by the dynamic movement primitive, and the probabilistic extraction algorithm is applied to extract primitives from the unlabeled continuous trajectory data. The operation primitives use the Gaussian mixture model to complete the training process based on the extraction and classification results of the trajectory primitives. The Gaussian mixture regression(GMR) algorithm is applied to predict the steering angle. The results show that the probabilistic extraction algorithm not only utilizes the correlation between representation and extraction but also uses the reasonable setting of the initial segmentation point, which improves the efficiency of the algorithm and makes the extracted motion primitives conform to specific driving assumptions. The proposed motion primitives can not only represent the driver's driving data with high precision but also have good generalization ability to deal with the desired position and time duration change when the motion primitives are regenerated. Finally, the motion primitive library describing the driving behavior under all conditions is established, and the applicability of the motion primitives to different driving situations is significantly improved.

Key words: intelligent vehicle, motion primitive, driving data, dynamic movement primitives(DMP), gaussian mixture model(GMM), gaussian mixture regression(GMR)

中图分类号: