本文为瑞典查尔姆斯理工大学（作者：Robert Hult、Reza Sadeghi Tabar）的信号系统硕士论文，共145页。
在过去的二十年中，高度自动化车辆系统的愿景吸引了学术界和工业界研究人员的兴趣。本文的工作集中在该领域的路径规划子集上，该子集解决了车辆在充满障碍物的环境中如何导航的问题。其目的在于研究和评估路径规划方法的现状，并提出在当今客车存在的资源约束下可行的、满足鲁棒性和乘客舒适性要求的解决方案。
由于一般性路径规划问题的复杂性，本课题研究范围仅限于在公路和乡村道路上的行驶，包括通过转弯处的入口和出口。提出了一种模块化的解决方案，将路径规划精简为一组简单轨迹集合的选择；这种选择基于过滤的复合代价函数，该函数考虑各种性能特点，例如与其它车辆之间的距离和舒适性等。对于高速公路和转弯机动，通过求解约束最优的控制问题，简化了车辆动力学的表示方法，从而求解得到简单的轨迹形式。结果表明，当使用粒子模型与特定的代价函数一起逼近动力学特征时，公路情况下的运动轨迹呈现为多项式形式。
然而，对于车辆转弯的情况，相应的约束最优控制问题是数值化的，通过假定控制问题的参数化形式，详细地推导了非线性规划的表达式。论证了初始猜测对数值求解的重要性，提出了一种利用前馈神经网络（FFNN）函数拟合生成算法起始点的新方法。证明了使用基于FFNN的高精度初始猜测的优越性，该算法优于其它文献中提出的方法。大量的仿真结果表明，本文提出的系统能够解决复杂交通场景下的规划问题。通过比较简单模型和非线性自行车模型的动力学特性，验证了简单模型的合理性，后者被迫遵循前者的设计路径。所推导的微分代数方程通过Dymola求解，并用于确定在什么条件下进行简化是有效的。最后，本文讨论了使用模块化方法解决路径规划问题的优缺点。
The vision of highly automated vehicle systems has for the last twodecades enticed researchers in both academia and industry. The work in thisthesis focuses on the path planning subset of the field, which addresses theproblem of how a vehicle should navigate through an obstacle filledenvironment. The purpose is to study and evaluate the state of the art of pathplanning approaches and propose a solution that is feasible under the resourceconstraints present in today’s passenger vehicles and that satisfies demands onrobustness and passenger comfort. Due to the complexity of the general pathplanning problem, the scope is limited to operation on highways and countryroads, including entries and exits through turns.A modular solution ispresented where the planning is reduced to selection among a discrete set ofsimple trajectories. The selection is based on a filtered compound costfunction that capture various performance aspects, e.g. proximity to othervehicles and comfort. For both highway and turning manoeuvres, the simpletrajectories are found by solving a constrained optimal control problemsubjected to simplified representation of the vehicle dynamics. It is shownthat the trajectories for the highway case takes the form of a polynomial whena particle model is used to approximate dynamics together with a specific costfunctional. For the turning case however, it is shown that the correspondingconstrained optimal control problem is of numerical nature and thereformulation to a non linear program by assuming a parametrized form of thecontrol is detailed. The importance of the initial guess to the numericalsolver is demonstrated and a novel way of generating these algorithmic startingpoints using function fitting via Feed Forward Neural Networks (FFNN) ispresented. The benefit of using the highly accurate FFNN based initial guess isdemonstrated, and its superiority over other methods presented in theliterature is shown. Results from extensive simulations are presented whichdemonstrates that the proposed system is capable to solve planning throughcomplex traffic scenarios. The model simplifications made are justified by acomparison between the dynamics of the simple model and a non linear bicycle model,where the latter is forced to follow the path of the former. The resultingDifferential Algebraic Equations are solved through Dymola and used todetermine under what conditions the simplification is valid. The thesis isconcluded with a discussion on the results touching on both the benefits anddisadvantages of using a modular approach to the path planning problem.
1 引言
2 项目背景
3 本文提出的路径规划系统
4 仿真结果
5 讨论
6 结论
附录A 方法调研
附录B 公路机动轨迹
附录C 代价分量
附录D 转弯轨迹
附录E 车辆建模
下载英文原文地址：
http://page5.dfpan.com/fs/blc9j28218294161dc1/
更多精彩文章请关注微信号：