【作者】 王萍；

【导师】 潘树国； 严后选；

【作者基本信息】 东南大学 ， 电子信息（专业学位）， 2025， 硕士

【摘要】 地下空间作为战略性资源,在保障安全、资源利用和促进发展等方面具有重要作用。随着狭窄地下环境中无人化、智能化的发展趋势,地下无人驾驶技术的研究与应用变得尤为关键。路径规划是无人驾驶车辆的关键技术之一,主要分为提供整体最优路径的全局路径规划,以及根据动态环境实时调整轨迹的局部轨迹规划。在结构规则、空间开阔的常规环境中,图搜索、采样类的路径规划算法已取得了良好的应用效果。然而,在结构复杂、空间狭窄的地下环境中,传统路径规划方法面临更加严峻的挑战:全局规划存在计算耗时长、生成路径质量差等问题,局部规划存在避障难度大、易陷入局部最优等问题,并且在动态场景中适应性差,轨迹稳定性难以保证。针对上述问题,本文提出了一种面向地下空间的无人车辆路径规划方法,通过构建全局-局部的分层规划框架,在全局路径规划层利用地图骨架约束的启发式图搜索方法提高了狭窄环境的路径搜索效率,在局部规划层引入基于有限状态机的动态窗口方法提升了自适应避障能力。论文的主要研究工作如下:(1)提出了一种基于地图骨架约束的启发式图搜索全局规划方法。针对传统图搜索算法在地下场景中耗时长、路径质量不佳的问题,本文提出了一种地图骨架约束的策略:利用图像处理的开运算与地图降采样方法对栅格地图进行预处理,有效优化地图边界并降低后续规划的计算开销;进一步采用Zhang-Suen细化算法提取通道中心骨架,为深度优先搜索(Depth-First Search,DFS)提供全局约束以减少无效遍历。此外,提出距离与角度启发式策略:对搜索的骨架点按目标接近程度与朝向合理性进行排序,引导算法快速朝正确方向扩展,进一步提升搜索效率。最后利用三次样条插值对初始路径进行平滑处理,确保路径的连续性与可行性。实验结果表明,在狭窄的地下场景中,本文的启发式图搜索方法相比于RRT、Skilled-RRT、ARRT-Connect、A*、骨架A*及双层A*算法,规划耗时减少超过54%,路径曲率较采样类的方法降低59%以上,在平滑性和可行性方面表现良好。(2)提出了一种基于行为预决策的自适应DWA(Dynamic Window Approach)局部轨迹规划方法。针对传统DWA算法在动态场景中适应性差、轨迹稳定性弱的问题,本文加入了有限状态机决策模块,根据环境信息预判车辆行为,并通过限制轨迹规划范围减少轨迹的不稳定性。同时,改进了DWA算法的轨迹采样过程与评价策略:在轨迹采样阶段,提出自适应预测时间调整和权重因子自适应策略,避免固定参数导致局部最优或规划失败的情况,提高算法的场景适应能力;在轨迹评估环节,加入全局路径贴近项,并优化安全性、速度及航向角评价项,引导算法优先选择平滑、贴合全局路径且安全高效的轨迹。实验结果表明,在地下狭窄场景中规避静态障碍物时,本文的自适应DWA方法相比于DWA、Fuzzy＿DWA和改进IDWA算法,计算耗时减少38%以上,平均行驶速度提升33%以上。此外,该方法具备超车、跟车、减速避让及优先通行的能力,能够动态调整轨迹与速度以适应复杂环境,展现出良好的环境适应性。(3)基于一套具备自主导航能力的无人车辆实验平台,开展路径规划方法的测试与验证。该平台的硬件系统由车辆底盘、计算单元及多种传感器组成,软件架构集成了自主定位、环境感知、路径规划、运动控制及数据交互等核心模块,实现了系统的高效协同运行。最后,提出了基于全局-局部分层的规划方法,并在狭长的类地下环境中开展实车测试,涵盖无障碍物、静态障碍物和动态障碍物三类场景。实验结果表明,搭载该规划方法的无人车辆具备全局规划能力,能够以0.2s的规划周期实时调整轨迹。车辆可以根据决策结果稳定避开静态障碍物,并在动态环境中具备超车、跟车、减速避让及优先通过的能力,验证了该方法的可行性与实用性。更多还原

【Abstract】 As a strategic resource,underground space plays a vital role in ensuring safety,optimizing resource utilization,and promoting development.With the trend of unmanned and intelligent development in narrow underground environments,the research and application of autonomous driving technology in such settings has become increasingly important.Path planning is a critical technology in autonomous driving,primarily consisting of global path planning that provides the overall optimal path,and local trajectory planning that adjusts the trajectory in real-time based on the dynamic environment.In structured and spacious conventional environments,path planning algorithms based on graph search and sampling have achieved favorable application results.However,in underground environments characterized by complex structures and narrow space,traditional methods face significantly greater challenges:global planning suffers from long computation times and poor path quality,while local planning faces challenges in obstacle avoidance and tends to fall into local optima.Moreover,in dynamic environments,it demonstrates limited adaptability and struggles to maintain trajectory stability.To address the aforementioned issues,this thesis proposes a path planning method for narrow underground scenarios.A hierarchical global-local planning framework is constructed:at the global planning level,a heuristic graph search method based on map skeleton constraints is employed to enhance path search efficiency in narrow scenarios;at the local planning level,a dynamic window approach based on finite state machines is introduced to improve adaptive obstacle avoidance capabilities.The main research work of the thesis is as follows:(1)A heuristic graph search global path planning method based on map skeleton constraints is proposed.To address the issues of long computation times and poor path quality in traditional graph search algorithms for underground scenarios,this thesis proposes a map skeleton constraint strategy:the grid map is preprocessed using morphological opening operations and map downsampling techniques,which effectively optimize map boundaries and reduce the computational cost of subsequent planning;furthermore,this thesis use the Zhang-Suen thinning algorithm to extract the central corridor skeleton,providing global constraints for Depth-First Search(DFS)and reducing invalid traversals.Additionally,the distance and angle heuristic strategy are proposed:skeleton points are ranked based on their proximity to the goal and the suitability of their heading,guiding the algorithm to expand efficiently in the desired direction and further enhancing search performance.Finally,the initial path is smoothed using cubic spline interpolation to ensure its continuity and feasibility.Experimental results indicate that in narrow underground scenarios,the heuristic graph search method in this thesis reduces planning time by over 54%compared to RRT,Skilled-RRT,ARRT-Connect,A*,Skeleton A*,and double-layer A*algorithms,achieves a reduction in path curvature of over 59%relative to sampling-based methods,and exhibits excellent smoothness and feasibility.(2)A dynamic window approach(DWA)local trajectory planning method based on behavior pre-decision is proposed.To address the poor adaptability and low trajectory stability of the traditional DWA in dynamic scenarios,this thesis introduces a finite state machine decision-making module,which predicts vehicle behavior based on environmental information and constrains the trajectory planning range to reduce trajectory instability.Furthermore,the DWA algorithm’s trajectory sampling process and evaluation strategy are improved:in the trajectory sampling phase,adaptive prediction time adjustment and weight factor adaptation strategies are proposed,avoiding local optima or planning failures caused by fixed parameters,and improving the algorithm’s adaptability to different scenarios;in the trajectory evaluation phase,a global path adherence term is introduced,and the evaluation metrics for safety,speed,and heading angle are optimized to guide the algorithm toward selecting smoother,globally-aligned,and safer trajectories with higher efficiency.Experimental results show that during static obstacle avoidance in narrow underground scenarios,the adaptive DWA method proposed in this thesis reduces computation time by more than 38%and increases average driving speed by more than 33%compared to DWA,Fuzzy＿DWA,and improved IDWA algorithms.Additionally,it has the ability to overtake or follow the preceding vehicle,decelerate and wait,and pass with priority,enabling dynamic trajectory and speed adjustments to adapt to complex environments,demonstrating strong environmental adaptability.(3)Based on a self-navigation capable unmanned vehicle experimental platform,this thesis conducts testing and validation of the proposed path planning method.The platform’s hardware system consists of a vehicle chassis,computing unit,and various sensors,while its software architecture integrates key modules including autonomous localization,environmental perception,path planning,motion control,and data communication,enabling efficient system coordination.Finally,a hierarchical global-local planning method is proposed and real vehicle testing is conducted in narrow and long environments,covering three scenarios:obstacle free,static obstacle,and dynamic obstacle.The experimental results show that the unmanned vehicle equipped with this planning method has global planning capability and can adjust its trajectory in real-time with a planning cycle of 0.2s.The vehicle can stably avoid static obstacles based on decision results and possesses the ability to overtake,follow,decelerate for avoidance,and give priority to passing in dynamic environments,verifying the feasibility and practicality of this method.更多还原