自主导航柑橘表型巡检机器人设计与试验

Design and Experimentation of Autonomous Navigated Citrus Phenotype Inspection Robot

摘要:

为了提高柑橘育苗的自动化水平，提出了一种适用于柑橘育苗的全自动表型巡检机器人。首先结合三维激光雷达与惯导信息对育苗环境进行SLAM建图，对得到的三维点云地图进行预处理与投影，得到适用于规划和导航的二维地图。然后，采用HDL_localization定位算法进行精准定位，并结合Dijkstra算法与TEB算法，实现在全局路径规划的同时优化局部路径，规划出理想的巡检路线，保障巡检的可靠性和安全性。在巡检过程中，工控机上运行的YOLO v8网络不断处理来自位于机器人两侧深度相机所拍摄的图像，识别出图像中的柑橘苗，计算得到株高，同时将这些数据实时上传至网络数据库。针对柑橘苗株高计算，提出并比较了3种不同的方法。试验结果证明，巡检机器人自动驾驶时的定位结果与从高精度RTK定位中获取的真值相比，平均定位误差为5.6 cm，最大定位误差为17.5 cm；使用最优的计算方法获取的柑橘苗高度与人工测量的真值相比，平均绝对误差为1.88 cm，最大绝对误差为7 cm，均方误差为5.93 cm2。

Abstract:

In order to improve the automation level of citrus nursery, a fully automated phenotype inspection robot suitable for citrus nursery was proposed. Firstly, SLAM mapping of the nursery environment was performed by combining 3D LiDAR and inertial guidance information, and the obtained 3D point cloud map was preprocessed and projected to obtain a 2D map suitable for planning and navigation. Then the HDL_localization positioning algorithm was used for accurate positioning, and combined with the Dijkstra algorithm and TEB algorithm, to achieve the optimization of local paths while global path planning, plan the ideal inspection route, and ensure the reliability and safety of inspection. During the inspection process, the YOLO v8 network running on the industrial computer continuously processed the images from the depth cameras on both sides of the robot, recognized the citrus seedlings in the images, calculated the plant height, and uploaded these data to the network database in real time. Three different methods were proposed and compared for citrus seedling plant height calculation. The experiments proved that the localization of the inspection robot on autopilot had an average localization error of 5.6 cm and a maximum localization error of 17.5 cm compared with the true value obtained from high-precision RTK localization, and the height of the citrus seedlings obtained by using the optimal computation method had an average absolute error of 1.88 cm, a maximum absolute error of 7 cm, and a mean-square error of 5.93 cm2 compared with the true value of the manual measurements.