王欢1  徐秀进1  王红军1  温志庆2

（1.华南农业大学工程学院，广东 广州 510642

2.季华实验室智能机器人工程研究中心，广东 佛山 528200）

摘要：针对采摘机器人自主行走导航过程中，难以准确定位其与果树之间的相对位置，难以准确估计果树树干姿态的问题，提出基于双目eye in hand系统的多角度树干位姿估计方法。利用YOLOv5深度学习方法与半全局块匹配算法识别树干并生成局部点云；利用半径滤波和体素滤波减少树干点云数据；利用闭环式手眼标定方法对双目eye in hand系统进行标定，并对同一树干多角度相机位置的点云数据进行拼接；利用随机抽样一致（RANSAC）算法与无约束最小二乘法估计并优化树干的位置和姿态，获取树干的圆柱体参数。通过对30幅标定板图像进行实验，闭环式手眼标定方法的平均欧式误差为3.717 7 mm；采用半径滤波和体素滤波可减少98.470%的点云数据；采用RANSAC算法、圆柱体估计算法拟合树干点云数据，得到圆柱体的半径r = 41.277 1 mm，RMAE = 2.571 56 mm，RRMSE = 2.989 36 mm；无约束最小二乘法优化后r = 39.402 8 mm，RMAE = 1.989 55 mm，RRMSE = 2.465 88 mm。该文通过对双目eye in hand系统进行标定，建立坐标系转换关系，多角度采集环境信息，准确定位机器人与果树之间的相对位置，估计果树树干的姿态。

关键词：eye in hand手眼标定；局部立体匹配；树干圆柱体估计；多角度点云拼接

中图分类号：S225; TP391.41         文献标志码：A        文章编号：1674-2605(2023)05-0006-08

DOI：10.3969/j.issn.1674-2605.2023.05.006

Multi-angle Tree Trunk Pose Estimation Method Based on Binocular       Eye in Hand System

WANG Huan1  XU Xiujin1  WANG Hongjun1  WEN Zhiqing2

(1.College of Engineering, South China Agricultural University, Guangzhou 510642, China

2.Intelligent Robot Engineering Research Center, Ji Hua Laboratory, Foshan 528200, China)

Abstract: In the process of autonomous walking and navigation for harvesting robots, it is difficult to accurately locate the relative position between them and the fruit tree, as well as to accurately estimate the posture of the fruit tree trunk. Based on the binocular eye in hand system, YOLOv5 deep learning method and classical stereo matching algorithm are applied to identify tree trunks and generate local point clouds. Radial filtering and voxel filtering are used to reduce the number of point clouds. A closed-loop hand eye calibration method is proposed to calibrate the system and concatenate point cloud data from multi angle camera positions on the same tree trunk, using the random sampling consensus (RANSAC) algorithm and unconstrained least squares method to estimate and optimize the trunk posture, and obtain the trunk cylinder parameters. Through repeated experiments on 30 sets of calibration board images, the average Euclidean error of the closed-loop hand eye calibration method is 3.717 7 mm; Reduce 98.470% of point cloud data by using radius filtering and voxel filtering; The RANSAC cylinder estimation algorithm fitting tree trunk point cloud data yields a cylinder radius of r = 41.277 1 mm, RMAE = 2.571 56 mm, RRMSE = 2.989 36 mm; After unconstrained least squares optimization, r = 39.402 8 mm, RMAE = 1.989 55 mm, RRMSE = 2.465 88 mm. The closed-loop hand eye calibration method proposed in this article calibrates the eye in hand system, establishes coordinate system conversion relationships, collects environmental information from multiple angles, effectively and accurately locates the relative position between the robot and the fruit tree, and estimates the posture of the fruit tree trunk.

Keywords: eye in hand calibration; local stereo matching; estimation of tree trunk cylinder; multi angle point cloud stitching