基于极限学习机的单木枝叶点云分类

Cloud Classification of Individual Trees Using Extreme Learning Machine

[目的]在林业测量中,利用树木的三维激光点云数据提取其结构信息以及模型拟合三维重建较为普遍,而林木枝叶三维点云分割则是林木参数提取以及三维重建的前提。[方法]利用扫描的单木三维点云数据,提取了单木点云的空间特征、反射强度、RGB色彩特征等多维特征,为提高分类的效率,通过随机森林算法按照其特征重要程度排序,除去冗杂的特征,保留RGB色彩、反射强度、法向分布特征作为分割依据。采用极限学习机(Extreme Learning Machine,ELM)对训练样本进行学习,并对原始数据进行枝叶点云的分类识别试验,分类正确率达到98.99%。[结果]在同等试验条件下,分别采用BP(Back Propatation)神经网络、LVQ(Learning Vector Quantization)神经网络、决策树(Decision Tree)、朴素贝叶斯(Naïve Bayes Classifier,NBC)、支持向量机(Support Vector Machine,SVM)进行分类,分类,其正确率分别为94.30%、91.26%、96.95%、85.67%、98.16%。[结论]试验结果表明极限学习机的分类效果较好。

[Objective] In forestry surveying, it was common to extract structural information from 3D laser point cloud data of trees and model fitting for 3D reconstruction. The segmentation of tree branch and leaf 3D point cloud was the premise of tree parameter extraction and 3D reconstruction.[Method] In order to improve the efficiency of classification, random forest algorithm was used to sort the points according to their importance degree, removing the complicated features, preserving the RGB color, reflective intensity and normal distribution characteristics as a basis for segmentation. Extreme Learning Machine (ELM) was used to learn the training samples, and the classification and recognition experiments on the original data were carried out. The classification accuracy rate reached 98.99%.[Result] Under the same experimental conditions, BP (Back Propation) neural network, LVQ (Learning Vector Quantization) neural network, Decision Tree, Nave Bayes Classifier (NBC) and Support Vector Machine (SVM) were used to classify, and the classification accuracies of which respectively were 94.30%, 91.26%, 96.95%, 85.67%, 98.16%.[Conclusion] The experimental results show that the classification effect of extreme learning machine is better.