Tree Segmentation And DBH Estimation From Point Clouds Based On Deep Learning

In the complex urban scenes,how to accurately and efficiently obtain various information of urban trees is of great significance to urban forest management and smart city construction.Lidar technology,which can quickly acquire high-resolution 3D point clouds,is one of the important technologies to obtain the information of these trees.People can use machine learning algorithms to intelligently extract measurable three-dimensional information from trees.This paper studies the tree segmentation of 3D point cloud and the estimation of tree stem diameter.It mainly includes the detection and accurate segmentation of trees in the large scene point cloud and the extraction of the individual tree stem and estimation of its diameter at breast height:(1)Detection and segmentation of tree in point cloud.Aiming at the huge amount of high-resolution three dimensional point cloud data collected by Lidar scanning system in complex urban scenes,a method for detecting canopy by local distribution of point clouds is proposed.First learn the local distribution of middle echo points through a convolution network,then use it to detect the canopy in other data and obtain the neighborhood of the canopy.There are a lot objects in the vicinity of the obtained canopy,such as the nearby buildings,the cars under the trees and so no.So,the Nonlocal network is constructed to further finely segment the point cloud around the canopy.These point clouds are classified into trees,bush and other objects.Finally,the whole framework was verified on the Semantic 3D reduced-8 set.The IoU of the tree point cloud segmentation reached 86.40%.In addition,the semantic segmentation network was tested on the Paris-Lille-3D dataset,and the average IoU was improved than several other networks.(2)Extraction of individual tree stem and estimation of DBH.In view of the fact that many urban trees may be connected into pieces and the trunks are concealed in the leaves,a voxel-based method and a PointNet-based method are proposed to extract individual tree stem from the tree point cloud.The results show that the PointNet-based method can extract trunks more accurately and completely.Then take a point cloud slice from the tree stem and project it onto a plane perpendicular to the stem.Finally,the projection point is denoised,and the circle fitting method is used to estimate the DBH of each tree.Good results were obtained with a root mean square error RMSE of 1.78 cm.