Detection And Classification Of Street Trees Based On Mobile Mapping LiDAR Point Clouds And Images

The inventory of urban road tree plays a vital role in the city’s environmental construction.With the development of technology,traditional methods of manually measuring trees have gradually been replaced by machine measurements.In this paper,the mobile LiDAR system is used to collect the point clouds and images of the trees on both sides of the urban road,and the data is used to estimate individual tree and classify the trees,aiming at serving the automated urban road tree survey.This paper is divided into three parts:individual tree extraction and parameters estimation based on the 3D point cloud,tree species classification based on deep learning,and the establishment of urban road tree inventory dataset.Firstly,based on Euclidean clustering and PC A feature extraction,an individual tree is extracted from the large-scale LiDAR point cloud data.Then,the parameters of an individual tree are been estimated automatically.The tree parameters include:height,DBH,canopy diameter,stem volume,and trunk posture.To estimate stem volume of the trees,because of the irregularity of the trunk,this paper proposes an adaptive segmentation method based on surface normal vector,which can automatically adjust the segmentation interval of trunks with different thicknesses in vertical segmentation.Compared with the fixed segment method of 0.05m,the accuracy has been improved by 0.02%.Since the ground truth values of tree parameters are difficult to obtain,this paper uses OnyxTree so:ftware to build tree models with different parameters,and uses model parameters to judge the accuracy of the proposed algorithm.Then,using the Faster R-CNN-based detection and classification model,the trees in the multi-view urban road images are detected and classified.The urban road trees of the dataset mainly include camphor trees and palm trees.The trained model achieved 85.05%mAP on the data set.After the tree detection,according to the image and point cloud mapping relationship,the information of the image and point clouds are fused together to improve the accuracy of individual tree point cloud extraction.Finally,based on the proposed methods of individual tree extraction,parameters’estimation and tree species classification,we build a dataset that provide 3D tree point clouds and 2D images.Considering that the quality of the point cloud will affect the accuracy of the estimated tree parameters,the paper evaluates the quality of mobile LiDAR point cloud and analyzes its possible impact on the estimation of the parameters by comparing the tree point cloud acquired by the mobile LiDAR with the static LiDAR,by which the trees are scanned in all-orientation.In addition,taking a section of urban road as an example,the location,type,image,point cloud and various parameters of trees are displayed in the association.The experimental results demonstrate our proposed method and the constructed dataset show the application prospects to the urban road tree inventory.