基于改进DGCNN的树木点云分割方法  

Point cloud segmentation method of trees based on improved DGCNN

作　　者：刘超[1] 卜鑫荣 刘慧[1] 杨官学[1] 沈跃[1] 徐婕 LIU Chao;BU Xinrong;LIU Hui;YANG Guanxue;SHEN Yue;XU Jie(School of Electrical and Information Engineering,Jiangsu University,Zhenjiang 212000,China)

机构地区：[1]江苏大学电气信息工程学院,江苏镇江212000

出　　处：《南京农业大学学报》2025年第1期240-248,共9页Journal of Nanjing Agricultural University

基　　金：国家自然科学基金项目(32171908)。

摘　　要：[目的]通过目标分割为果园喷雾机提供树木的表征信息,使喷雾机能够实现精准喷雾。在分割过程中,对苗圃中的树冠、树干等不同部位进行分割,可以帮助喷雾机对喷雾部分对靶,在果园或苗圃景观中实现自动导航以及精准喷药等操作。与图片相比,点云能够更好地表征树木的三维结构并且受照明条件影响小,因此针对点云树木设计分割算法更适合应用在果园、苗圃等室外环境作业的农业机械。[方法]本文基于DGCNN提出了一种分割精度准确、参数量小的树木点云分割网络——TSNet,它可以很容易被部署在果园喷雾机上。该网络主要具有以下特点:1)该网络是基于DGCNN改进的,可以更好实现点云分割任务;2)网络引入了连续递归门控卷积模块(g^(n)Conv),可以提高树木分割的准确率;3)为避免全局信息损失并增加信息传递效率,我们设计了权重通道用于特征传递。[结果]TSNet分割树木的mIoU达到90.08%,模型大小为0.72 M,优于PointNet、PointNet++、DGCNN、CurveNet、PointMLP和D-PointNet++等常用的点云分割算法。[结论]TSNet能够为苗圃树木检测识别和农业机器人作业提供更准确的感知信息。[Objectives]The object segmentation is a fundamental vision task for orchard sprayers to achieve precision spraying of trees by providing their phenotypic characteristics. In the kind of task,the trees planting in the nursery can be segmented with different parts including crowns,trunks and so on,which can provide perceptual information to the sprayer and help them realize the accurate spraying or navigation operation in the orchard or nursery scenery without manual intervention. Comparing to 2D images,point clouds are more suitable for being employed in some outside environments like orchards and nurseries,as point clouds enable the sprayer to acquire 3D structures of trees and are independent from illumination. [Methods]In this study,we introduced TSNet,a point cloud segmentation network with smaller module size that can be easily deployed on orchard sprayer. The network mainly has the following advantages:1)The network can realize the segmentation task of tree point clouds,which improves based on DGCNN. 2)The introduction of g^(n)Conv based continuous recursive gate convolution(gConv)operation is built to improve the effectiveness of segmentation in trees. 3)The weight channel is designed to avoid loss of global features caused by process of deep mining. [Results]The mIoU value of TSNet tree segmentation reached 90.08%,and the number of model size was 0.72 M,which was better than the commonly used point cloud segmentation algorithms:PointNet,PointNet++,DGCNN,CurveNet,PointMLP,and D-PointNet++. [Conclusions]The proposed method can provide more accurate perceptual information for nursery tree detection and identification and agricultural robot operation.

关 键 词：点云 树木分割 深度学习 精准喷雾 果园喷雾机 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]