基于3D点云分析的果园行间穿梭机器人路径规划方法

Path Planning Method for Inter-row Shuttle in Densely Planted Orchards

摘要:

针对现有果园导航方法易受冠层密度、光照条件、种植不规整、地面不平整等条件影响，导致用于实现自主导航的树行方向估计方法与行尾识别方法稳定性低的问题，本文提出基于3D点云分析的果园行间穿梭路径规划方法，该路径规划方法包含树行识别定位方法、场景识别方法、路径规划方法，用于密植果园机器人行间自主穿梭的导航系统。首先，设计了基于点云语义分割网络的果树树干点云提取方法，实现了树行的识别与定位；其次，设计了基于卷积神经网络的位置场景识别方法，实现了行头等位置的场景识别；最后，设计了基于有限状态机的行间行进策略管理方法与基于RS曲线的行间路径规划方法，实现了果园多行连续行走。基于本文方法的树干点云分割平均交并比为88.3%，果树平均定位误差为2.04%(x方向)、1.54%(y方向)，树行方向估计平均误差为1.11°，行尾识别正确率为96%，行内中线行走平均偏差为0.08m。实验结果表明，本文所提出路径规划方法能够满足果园环境下树行定位与位置场景识别准确性要求，有效规划行间穿梭路径，为果园激光自主导航提供有效参考。

Abstract:

Addressing the issues with existing orchard navigation methods, which are susceptible to canopy density, lighting conditions, irregular planting, and uneven ground, leading to low stability in tree row direction estimation and row end identification methods used for autonomous navigation, a 3D point cloud-based method for pathway planning was introduced. The method comprised three primary components, a tree row identification technique, a scene recognition method, and a pathway planning strategy. These formed a system that enabled robots to autonomously shuttle between densely planted rows. Initially, a trunk point cloud extraction method using a semantic segmentation network was developed for tree row identification and positioning. Subsequently, a convolutional neural network-based method was established for location scene recognition to identify various scenarios within rows. Finally, an inter-row movement strategy managed by a finite state machine and a pathway planning method based on RS curves were designed for continuous walking through multiple rows. The trunk point cloud segmentation achieved an average IoU of 88.3%, with tree localization errors of 2.04% in the x-direction and 1.54% in the y-direction. The average error in tree row direction estimation was 1.11°.The row-end recognition accuracy reached 96%,and the average deviation of in-line centerline walking was 0.08m. These results showed that the proposed methods met the accuracy requirements for tree row positioning and scene recognition in outdoor orchards, which effectively planed pathways between rows and served as a reliable reference for autonomous laser-guided navigation.