柔性障碍物富集环境中的三维自主导航研究

Autonomous navigation is an edge and hot research topic in robotics, which has a decisive role in the application-field expansion of robots. However, the existing algorithms of autonomous navigation do not pay enough attention on the rigidity and deformability of the obstacles, leading to the simplified assumption that all the obstacles are rigid. Under such an assumption, the navigation algorithm can only choose to sacrifice the optimality by avoiding all the obstacles, while some soft and deformable ones can be tackled in a touching manner. As a result, the navigation performances in some application scenes, such as the woods/farm with rich soft branches and the urban with lots of flags/curtains, are greatly decreased. To solve the problem, this project introduces the concept of "soft-and-deformable obstacles" and tries to build a complete solution framework for the autonomous navigation in the environments with rich soft-and-deformable obstacles. The framework consists of three parts: the detecting and recognition strategy, the three-dimensional expression method, and the navigation algorithm. To discriminate the soft-and-deformable obstacles from the rigid obstacles, we will use the RGB-D (red, green, blue and depth) sensor and the laser range founder in combination. The RGB-D sensor will be employed to obtain the color and depth features of the obstacles, and the laser range founder will be employed to obtain the beam reflection energy information of the obstacles. Referring to a priori knowledge library build before hand, soft-and-deformable obstacles can be recognized with a high success rate. To generate the proper descriptions for the soft-and-deformable obstacles, a new pass cost function will be presented firstly, concerning both the collision performances and the levels of protection. Then, the multiple-volume-occupancy-grids (MVOG) map will be improved to be capable of storing the pass cost information. To navigate the robot in the environment with rich soft-and-deformable obstacles, the 2.5D angle potential field algorithm will be upgraded to be able to handle 3D cases, and will be endowed with the capability of tackling the obstacles in a touching manner. The proposed navigation framework enriches the definition of obstacle with the concept of soft-and-deformable obstacle, changes the embarrassing situation that obstacles can be only avoided with no contact, and brings more optimal performances for the navigation in the environment with rich soft-and-deformable obstacles. In conclusion, the research of the new navigation framework is a significant supplement for the theories of environmental perception and decision making in robotics, and is good for the rapid expansion of robotic application fields, which definitely has important scientific values.

机器人的自主导航能力是影响其应用领域拓展进程的关键因素。现有的自主导航理论对障碍物的材质特性关注度不够，忽略了质地柔软、可接触式通过的障碍物在提升导航优化性方面的重要作用，极大制约了机器人在树林、农田等场景中的应用表现。本项目引入柔性障碍物的概念，探索建立一套柔性障碍物富集环境下的移动机器人自主导航方法体系：组合使用"彩色测距"和表面反射能量分析技术，实现对柔性障碍物的有效探测与识别；提出兼顾障碍物碰撞特性、受保护程度的通过代价函数构建方法，实现柔性障碍物信息在三维环境模型中的合理表述；引入"接触式处理模式"，形成柔性障碍物富集环境中的优化自主导航算法。项目提出的方法体系拓展了障碍物的定义，改变了单一的"躲避式"导航思路，提高了导航算法在柔性障碍物富集环境下的优化性能，是对机器人环境感知和优化决策能力的有益理论补充，有利于迅速扩大机器人的应用场景及应用方式，具有重要的科学意义和应用价值。

地面移动机器人自主导航是机器人领域的热点研究问题，吸引了大量的研究者开展相关工作。虽然地面移动机器人自主导航研究已经取得了可喜的成果，但是现有导航算法中缺乏对柔性障碍物的识别和处理方法。针对这一问题，本项目展开柔性障碍物识别、模型表述和导航算法处理三部分的研究。在实际工作中，项目团队结合承担单位的研究领域特点和现有条件，选择了农业场景作为典型柔性障碍物富集环境，重点研究针对温室、大田作物/环境的识别算法、环境模型表述方法以及相应的导航方法，具体开展了农业移动机器人平台搭建、基于机器视觉的农业机器人环境感知算法研究、基于激光测距和惯性导航数据的农业机器人自主定位与导航算法研究等工作。农业移动机器人平台搭建方面，先后搭建了3台农业移动机器人平台样机，其中2台用于温室内部，1台用于大田环境中，每台样机均进行了机构设计、动力控制和系统集成等工作，并以样机为实验平台开展了农业机器人自主定位和导航算法的研究与实验；基于机器视觉的农业机器人环境感知算法研究方面，研究了基于单目图像处理的大田作物行提取算法、基于彩色测距（RGB-D）信息的温室行间过道边界提取算法和基于彩色测距信息的温室成簇番茄果实检测算法；基于激光测距和惯性导航数据的农业机器人自主定位与导航算法方面，研究了基于激光测距仪的温室道路跟踪导航算法、基于编码器数据的BP神经网络定位算法。项目组共发表论文5篇，录用论文1篇，在审论文1篇，其中EI检索论文3篇，，EI录用待检索1篇，硕士学位论文1篇。基于以上研究工作，项目建立了农业场景中柔性障碍物富集条件下的移动机器人自主导航方法体系雏形，为将来深入开展相关方向的研究工作奠定了基础。