| 基于改进人工势场的无人驾驶动态规划算法研究 |
| |
| 引用本文： | 罗玉涛,石紫娴,梁伟强.基于改进人工势场的无人驾驶动态规划算法研究[J].中国公路学报,2022,35(12):279-292. |
| |
| 作者姓名： | 罗玉涛 石紫娴 梁伟强 |
| |
| 作者单位： | 1. 华南理工大学 机械与汽车工程学院, 广东 广州 510640;2. 广州汽车集团股份有限公司汽车工程研究院, 广东 广州 510640 |
| |
| 基金项目： | 广东省自然科学基金项目(2016B01032001) |
| |
| 摘 要： | 无人驾驶决策算法可以分为端到端的决策算法与分层式决策算法，分层式算法由于可解释性强、鲁棒性高而被大多数主机厂采用。规划模块是分层式决策算法中的核心模块，它承接感知与地图模块的信息并输出驾驶轨迹或动作，而人工势场法由于规划效率高、信息提取能力强，被越来越多地应用于无人驾驶决策规划领域。但现阶段的人工势场存在未考虑目的地因素或建立目的地单点引力场导致远距离引力过大、方向错误的问题，无法应对复杂交通环境。针对这些问题，提出一种无人驾驶“行车意图-风险复合场”(Driving Intention & Risk Field, IRF)，根据目的地、车辆、道路边界等要素各自的特点分别建模，并以势场的形式统一在IRF中。创建考虑全局规划的全局引力场，将全局规划路径离散成等距离的路径点，并动态选取感兴趣范围内的路径点进行全局引力场的构建。为了验证模型的性能，搭建IRF-SAC动态规划算法平台，并在CARLA仿真环境分别设置高速公路场景、十字路口场景和环岛场景。研究结果表明:相比于NF-SAC和FSM，IRF-SAC算法在安全性、舒适性、通行效率上均有显著提升；在高速公路场景下，IRF-SAC显示出较强的路径跟踪精度和鲁棒性，最大位移偏差相对于NF-SAC和FSM算法分别下降了44.8%、70.2%；在十字路口场景下，与NF-SAC及FSM算法相比，平均危险系数分别降低12.0%、20.6%，纵向加速度均方根分别降低13.2%、44.9%，行驶时长相较于FSM算法减少了39.2%；在环岛场景下，与NF-SAC及FSM算法相比，平均危险系数分别降低了31.7%、52.9%，纵向加速度均方根分别降低了27.0%、19.0%。
|
| 关 键 词： | 汽车工程 自动驾驶 人工势场 动态规划 智能决策系统 强化学习 |
| 收稿时间： | 2021-10-22 |
Autonomous Driving Dynamic-programming Algorithm Based on Improved Artificial Potential Field |
| |
| LUO Yu-tao,SHI Zi-xian,LIANG Wei-qiang.Autonomous Driving Dynamic-programming Algorithm Based on Improved Artificial Potential Field[J].China Journal of Highway and Transport,2022,35(12):279-292. |
| |
| Authors: | LUO Yu-tao SHI Zi-xian LIANG Wei-qiang |
| |
| Affiliation: | 1. School of Mechanical & Automotive Engineering, South China University of Technology, Guangzhou 510640, Guangdong, China;2. GAC Automotive Engineering Institute, Guangzhou 510640, Guangdong, China |
| |
| Abstract: | Decision-making algorithms for autonomous driving can be divided into end-to-end and sequential planning algorithms. Sequential algorithms are adopted by most OEMs because of their interpretability and robustness. Planning module is the core of the sequential algorithm. It receives information from a perception module and high-definition map and outputs the driving trajectories or actions. Artificial potential field (APF) method, which is widely used in planning algorithms for autonomous driving, is becoming increasingly popular owing to its excellent planning efficiency and information extraction capability. However, APF does not consider destination factor, and single-point destination gravitational field causes a large force, resulting in incorrect directions for long-distance cases, and cannot cope with complex traffic environments. In response to these problems, this study proposed “Driving Intention & Risk Field” (IRF) to model traffic factors including destination, vehicles, and road boundaries and consider their characteristics separately and then in combination. A global gravitational field considering the global route was created, and a global planned path was discretized into equidistant path points. The path points within the range of the interest area were dynamically selected to construct a global gravitational field. To verify the performance of the IRF, an IRF-SAC decision-planning algorithm platform was built, and highway, urban crossroad, and roundabout scenes were set in a CARLA simulation environment. The research results show that compared with NF-SAC and FSM, the IRF-SAC algorithm significantly improves safety, comfort, and vehicle-passing efficiency. In the highway scenario, IRF-SAC achieves high accuracy and robustness in path tracking, and the maximum displacement errors are reduced by 44.8% and 70.2% compared with the FSM and NF-SAC algorithms, respectively. In the crossroad scenario, the average risk coefficients are reduced by 12.0% and 20.6%, and root mean squares of the longitudinal acceleration are reduced by 13.2% and 44.9%, compared with the NF-SAC and FSM algorithms, respectively. Moreover, the driving time is reduced by 39.2% compared with the FSM algorithm. In the roundabout scenario, the average risk coefficient is reduced by 31.7% and 52.9%, and the root mean square of the longitudinal acceleration is reduced by 27.0% and 19.0%, compared with the NF-SAC and FSM algorithms, respectively. |
| |
| Keywords: | automotive engineering autonomous driving artificial potential field dynamic programming intelligent decision-making system reinforcement learning |
|
| 点击此处可从《中国公路学报》浏览原始摘要信息 |
|
点击此处可从《中国公路学报》下载全文 |
|
