Research On Individual Tree Segmentation And Structural Parameter Extraction Based On Laser Point Cloud Data

The collection of street tree information is an important part of urban ecological construction work.The handheld LiDAR is used in a bottom-up collection method to obtain point cloud data of target objects.The handheld LiDAR has strong penetration ability and the data collected by it has high accuracy.It can quickly and accurately obtain forest structure information,especially the crown and its lower branches.The segmentation of street tree LiDAR point clouds is the basis for obtaining the parameters of individual tree structure.The current individual tree segmentation methods mainly focus on airborne LiDAR point clouds,using elevation information to form a grid map for segmentation,or using canopy vertices as seed points for clustering.The disadvantage is that it is easy to ignore the information of understory branches.However,side-view LiDAR(vehicular LiDAR and handheld LiDAR,etc.)can comprehensively obtain roadside tree canopy and understory information.The lack of data and the confusion of tree spatial structure caused by data occlusion make the structure of the tree point clouds more complex,making it more difficult to achieve individual tree segmentation and parameter extraction.To address the problem of missing information under the forest and complex tree point cloud structures mentioned above,relevant research has been conducted.The main tasks are as follows:(1)Classify point clouds in street tree scenes based on spatial geometric features,obtain the part of trees,and prepare for subsequent individual tree segmentation.To classify vegetation more accurately,this paper studies point cloud classification based on geometric feature extraction and Fisher algorithm,and proposes two new 3D spatial features.The point cloud data collected by ground-based LiDAR and handheld LiDAR are used to classify vegetation,verifying the robustness of the classification algorithm.In the data collected by ground-based LiDAR,the weights of two new features(Area feature and Pointing feature)calculated by Fisher algorithm are 7.25 and 5.78,and the weight of Area feature is only second to the feature with the highest weight which is an eigenvalue of the point cloud covariance matrix.The accuracy of the classification using original features is 99.15%.After adding two new features,the accuracy is improved by0.75%,and the classification effect at the intersection of tree trunks,ground,and shrubs is significant.The results show that the proposed new features have high weights and can effectively improve the accuracy of vegetation classification.The classification effect of data collected by handheld LiDAR is also good,with an accuracy of 99.74% after using new features.(2)Propose an individual tree segmentation method based on dimension transformation.After separating the part of trees based on the geometric features of the point clouds,the optimal projection of the 3D point clouds on the 2D plane is calculated.The image segmentation algorithms are used to extract the individual tree edge of the 2D image,and the result is the corresponding 3D point cloud contour matched based on the individual tree edge pixel points in the 2D image.The LiDAR data of urban street trees were tested using this method,and the accuracy,recall,and quality of the proposed individual tree segmentation method are 91.67%,85.33%,and 79.19%,which are superior to the CHM-based method by 2.70%,6.19%,and7.12%,respectively.(3)Calculate the structural parameters of an individual tree.Using the results of individual tree segmentation in this study,it is possible to directly calculate the height and diameter at breast height of a tree.The tree height depends on the difference between the maximum and minimum of the Z coordinate of the result of segmentation.The diameter at breast height is the average diameter at breast height at 1.25 m,1.3m,and 1.35 m from the tree root.The extraction of crown diameter is based on an improved Graham method,which simplifies the projection image of the calculated convex hull,thereby avoiding the operation of invalid points inside the image and improving the efficiency of the algorithm.To quantitatively analyze the effect of parameter extraction,the results of manual measurement and algorithm extraction are fitted by linear regression.The regression evaluation index of tree height: R2=0.920,RMSE=0.261.The regression evaluation index of crown diameter:R2=0.914,RMSE=0.156.The regression evaluation index of diameter at breast height: R2=0.985,RMSE=0.024.The experiment shows that the algorithm extracted values of the three parameters has little difference from the manually measured values,and the accuracy of diameter at breast height is higher than that of tree height and crown diameter.This result verifies the effectiveness of the parameter extraction method proposed in this paper and the strong ability of handheld LiDAR to obtain understory information.