| LiDAR data is collected by laser scanners and has a wide range of applications in the field of forest information management.The three-dimensional structure of point cloud data can accurately reflect the structural information of trees,and can accurately obtain the location information,height information and tree crown area information.Although point cloud data has the advantage of spatial information that cannot be reflected by aerial images,the occlusion problem between trees may lead to incomplete tree structure information,and at the same time,obtaining high-density point cloud data requires higher cost.Most of the existing tree detection methods based on LiDAR data have the problem of not being able to detect the underlying trees and distinguish the trees and other objects in the scene.In order to solve the problems in the tree detection method,this paper carries out the research content from the following three points:1.Aiming at the difficulty in extracting lower trees in the process of multi-layer tree detection,this paper proposes a multi-layered tree crown extraction algorithm using graph-based segmentation.Firstly,the method uses the pit-free algorithm to rasterize the LiDAR data,and uses the graph-based segmentation algorithm to extract the overstory tree information in the rasterized canopy height model.Then,the three-dimensional structure analysis is performed on each of the divided tree areas,and it is judged whether there is a understory tree,and if there is a hierarchical structure in the current area,the corresponding LiDAR data is extracted.Finally,all the extracted point cloud data are combined and segmented to extract the understory tree information.In this paper,six experimental areas are selected as test areas from the two data sets.The experimental results show that the method can accurately extract the understory trees under the premise of ensuring the results of the overstory trees detection.2.Most of the existing detection methods are analyzed in pure tree scenes,so misunderstandings may occur when buildings or other non-tree objects appear in the scene.In order to solve this problem,this paper proposes a tree detection algorithm for complex scenes based on PointCNN.The method randomly extracts a large number of samples in the data set to train the PointCNN.In the detection process,the seed point is obtained by the local maxima with variable window sizes,and the different scales of sample is extracted using the seed point,and finally the sample is input into the PointCNN for detection.In this paper,the experimental data is divided into pure tree scenes and mixed complex scenes,and six experimental areas are selected to verify the method.The experimental results show that the method has better detection results in the pure tree scene,and can more accurately distinguish the trees and other objects in the mixed complex scene.3.Based on the above two methods of tree detection,this paper designs and implements a tree detection system based on LiDAR data.The system performs tree detection on the point cloud data input by the user,thereby extracting location information of the trees and tree height information.The system can perform tree detection through multi-layered tree crown extraction algorithm using graph-based segmentation and tree detection algorithm for complex scenes based on PointCNN. |
