Abstract: Inter‑line mechanical autonomous navigation is helpful to improve fruit production efficiency and reduce labor cost. Trees are natural landmarks for navigating between lines and can provide cues for robots. In this paper, combining deep learning and least square method, a navigation line extraction method based on machine vision for inter‑line navigation scene is proposed. Firstly, the tree trunk image was collected in the actual environment, and the image was flipped and clipped to expand the tree trunk data set. Secondly, the YOLOv5 network model was constructed, and based on this model, the interline trunk identification was carried out. The method of replacing the middle point of the root with the middle point under the identification box was proposed, so as to determine the positioning basis point of the tree line fitting. Finally, the least square method was used to fit the single tree line and center line of tree line in orchard. The experimental results show that the average recognition accuracy of the YOLOv5 network detection model is 85.5%. In the proposed root point replacement positioning method, the average error of the distance between the midpoint of the bottom of the identification box, the linear pixel distance between the positioning base point and the actual root midpoint, is 5.1 pixels, and the selection error of the positioning base point is within the reliable range. The average lateral deviation of the navigation line in the center of the tree line is 5.8 pixels, which meets the requirements of inter‑line navigation.

Key words: orchard navigation, deep learning, YOLOv5, tree trunk recognition, root point replacement, linear fitting

摘要： 行间机械自主导航有助于提高果品生产效率，降低人工成本。树木是行间导航的天然地标，可以为机器人提供导航信息。结合深度学习和最小二乘法，提出一种基于机器视觉的行间导航场景的导航线提取方法。首先，收集实际环境下的树干图像，并对图像进行翻转、裁剪等操作扩充树干数据集；其次，构建YOLOv5网络模型，并基于该模型来对行间树干进行识别，提出利用识别框下边中点替换根部中点的方法，以此来确定树行拟合的定位基点；最后，基于最小二乘法完成果园单侧树行线和树行中心线的拟合。试验结果表明，所构建的YOLOv5网络检测模型对树干的平均识别正确率为85.5%。所提出的根点替换定位法的直线像素距离平均误差为5.1像素，树行中心导航线的平均横向偏差为5.8像素，符合行间导航的要求。

关键词: 果园导航, 深度学习, YOLOv5, 树干识别, 根点替换, 直线拟合