Abstract: Aiming at the problem that manual measurement of nursery canopy parameters was timeconsuming and laborintensive, and fruit tree canopy parameters  could not be quickly extracted, a tree canopy information extraction method based on environmental point cloud recognition algorithm was proposed in this paper. Firstly, the LiDARIMU tightly coupled odometer was used for point cloud correction and feature point extraction, and the rotation constraint was used to solve the Zaxis migration problem in the construction map to complete the environment reconstruction of the measurement area. After the point cloud map was transferred to the background workstation, European clustering and 3DFVDNNs algorithm were used to segment and identify the tree point cloud. Finally, after finding the first main branch, the canopy volume was modeled by cubic voxel method, and the canopy area parameters were extracted by twodimensional raster method. The test showed that the mapping algorithm adopted in this paper could reconstruct the complete orchard environment with high accuracy. The Bet value of PR curve obtained by the nursery recognition method based on DNN deep learning classifier was 0.0641 and 0.0999 higher than that obtained by SVM and RF classifier. In addition, R2 and RMSE of crown volume and area were 0.74677 and 0.6978, 0.09754 and 0.07677, respectively. The results showed that the canopy parameters measured by the proposed algorithm were strongly correlated with the manual measurements, which provided important support for the fine management of orchards.

Key words:  , mobile robot, point cloud environment map, deep learning, point cloud recognition, canopy parameters

摘要： 针对人工测量苗圃冠层参数费时费力，无法快速提取果树冠层参数的问题，提出一种基于环境点云识别算法的树木冠层信息提取方法。首先利用LiDAR-IMU紧耦合里程计进行点云矫正和特征点提取，在建图中使用旋转约束解决Z轴偏移问题，完成测量区域的环境重建；将点云地图传输至后台工作站后，使用欧式聚类和3D-FV-DNNs算法对树木点云进行分割与识别；最后在找到第一主枝后利用立方体素法实现冠层体积建模，采用二维栅格法提取冠层面积参数。试验表明：本文采用的建图算法能较高精度地重建完整果园环境，基于DNN深度学习分类器的苗圃识别方法获取的PR曲线的Bet值比SVM与RF分类器所获取的数值高出0.064 1与0.099 9，此外树冠体积与面积的R2与RMSE分别为0.746 77、0.697 8以及0.097 54、0.076 77。表明本文算法测得的冠层参数与人工测量值有强相关性，为果园精细化管理提供重要支撑。

关键词: 移动机器人, 点云环境地图, 深度学习, 点云识别, 树冠参数