Search

CN-121999215-A - Single-wood point cloud instance segmentation method, device, equipment and storage medium

CN121999215ACN 121999215 ACN121999215 ACN 121999215ACN-121999215-A

Abstract

The invention discloses a single-tree point cloud instance segmentation method, device, equipment and storage medium, which comprises a scene semantic segmentation step, a Z-axis standardization step, a trunk height segment interception and three-dimensional to two-dimensional step, a section detection and root seed generation step and a three-dimensional growth type aggregation segmentation step, wherein stable single-tree instance segmentation under a complex topography and tree adjacent overlapping scene is realized by constructing a closed loop flow combination of semantic segmentation prefiltering, ground reference standardization, trunk height segment two-dimensional enhancement, section detection generation root seed and three-dimensional aggregation segmentation, and reliable input data is provided for subsequent tree parameter calculation and greening asset statistics.

Inventors

  • LI HANLONG
  • ZHU QIUYUE
  • LI SHIHONG
  • CHEN JIAN
  • MENG LIANGHUA

Assignees

  • 星景科技有限公司

Dates

Publication Date
20260508
Application Date
20260119

Claims (10)

  1. 1. The single wood point cloud instance segmentation method is characterized by comprising the following steps of: A scene semantic segmentation step, namely performing semantic segmentation on the three-dimensional point cloud of the landscaping scene to obtain a tree point cloud and a ground point cloud; a Z-axis standardization step, namely estimating local ground heights by taking the ground point cloud as a reference, and executing Z-axis standardization on the tree point cloud to ensure that the ground heights at different positions are uniformly mapped to 0; A trunk height section intercepting and three-dimensional converting step, namely intercepting trunk height section point cloud in standardized tree point cloud, and generating a two-dimensional thermodynamic diagram from three-dimensional projection of the trunk height section point cloud; Performing cross section detection and root seed generation, namely performing binarization and circle detection on the two-dimensional thermodynamic diagram to obtain trunk cross section circle center coordinates and radius information, and mapping the circle information back to a three-dimensional root seed point; and a three-dimensional growth type aggregation segmentation step, wherein three-dimensional growth type aggregation segmentation is carried out on the tree point cloud based on the root seed points and trunk section information, and a single-tree instance segmentation result is output.
  2. 2. The Shan Mudian cloud instance segmentation method as claimed in claim 1, wherein the semantic segmentation uses spatial coordinates xyz and normal vector information of points as input features, and RGB information and intensity information are not used in the semantic segmentation stage, and the semantic segmentation divides the point cloud into semantic categories including at least trees, roads, grasslands, pedestrians, automobiles, buildings and facilities.
  3. 3. The Shan Mudian cloud instance segmentation method as set forth in claim 1 or 2, wherein the pre-processing grid scale gridsize of the semantic segmentation stage is set to 0.02, and the semantic segmentation network adopts a PointTransformer deep learning network, wherein the PointTransformer deep learning network adds a scale self-adaptive multi-neighborhood feature fusion module, and introduces a boundary sensitive feature enhancement module at a classification head.
  4. 4. The Shan Mudian cloud example segmentation method according to claim 1 is characterized in that the ground point cloud is composed of road point cloud and grass point cloud obtained through semantic segmentation, local ground height is estimated through sliding window grid or local plane fitting in the Z-axis standardization step, wherein the sliding window size is 1m and is used as a step length at the same time, a ground height cutoff threshold value is set to be 0.25 and used for adapting to a slope scene with the step height and the gradient not greater than 27 degrees, and when a tree point corresponds to a grid without the ground height, an effective grid is found in a surrounding neighborhood range of the tree point, and an average value is taken as a ground height reference.
  5. 5. The Shan Mudian cloud example segmentation method according to claim 1 is characterized in that in the trunk height segment intercepting and three-dimensional converting step, point clouds with the height ranging from 0.2m to 1m from the ground are selected as trunk height segment point clouds, a cross section of 0.8m is selected for circle detection, DBSCAN density clustering is adopted on the trunk height segment point clouds, connected clusters with the largest density are reserved for filtering scattered points, the trunk height segment point clouds are projected to a horizontal plane to generate a two-dimensional thermodynamic diagram, and Gaussian filtering is carried out on the two-dimensional thermodynamic diagram and then normalization is carried out on the two-dimensional thermodynamic diagram to 0-255 gray level diagrams.
  6. 6. The Shan Mudian cloud instance segmentation method according to claim 1, wherein the binarization adopts an OTSU automatic thresholding method, the circle detection adopts hough circle transformation, the parameters are dp=0.8, minDist =20, param1=100, param2=13, minRadius =4 and maxRadius =32, each detected circle corresponds to a trunk section, and a two-dimensional circle center and a radius are mapped back to a three-dimensional coordinate system to form a root seed point, wherein the Z value of the root seed point takes the average height of the trunk section corresponding to the circle.
  7. 7. The Shan Mudian cloud instance segmentation method according to claim 1, wherein in the three-dimensional growing type aggregation segmentation step, when original point cloud data contains intensity information, linear values are calculated based on intensity and combined with PCA analysis to extract twigs, wherein the PCA neighborhood radius is set to 0.15, the wooden parts of a single tree are subjected to voxelization and three-dimensional refinement to obtain a central skeleton, the skeletons are subjected to breakpoint connection, skeleton end points with the degree of 1 are found, and connection is established when the distance between the two end points is smaller than d_max and the direction included angle is smaller than θ_max, wherein d_max=0.25 m and θ_max=60 degrees.
  8. 8. A single-wood point cloud instance segmentation apparatus, comprising: The semantic segmentation module is used for performing semantic segmentation on the three-dimensional point cloud of the landscaping scene to obtain a tree point cloud and a ground point cloud; the Z-axis standardization module is used for estimating local ground heights by taking the ground point cloud as a reference, and executing Z-axis standardization on the tree point cloud to enable the ground heights at different positions to be uniformly mapped to 0; The three-dimensional-to-two-dimensional module is used for intercepting trunk height section point cloud from the standardized tree point cloud, and generating a two-dimensional thermodynamic diagram from three-dimensional projection of the trunk height section point cloud; the section detection and root seed generation module is used for performing binarization and circle detection on the two-dimensional thermodynamic diagram to obtain trunk section circle center coordinates and radius information, and mapping the circle information back to a three-dimensional root seed point; and the three-dimensional growth type aggregation segmentation module is used for executing three-dimensional growth type aggregation segmentation on the tree point cloud based on the root seed point and trunk section information and outputting a single-tree instance segmentation result.
  9. 9. An electronic device comprising a processor and a memory, the memory storing a computer program, the processor implementing the Shan Mudian cloud instance partitioning method of any one of claims 1 to 7 when executing the computer program.
  10. 10. A computer readable storage medium, characterized in that a computer program is stored, which when executed by a processor implements the Shan Mudian cloud instance segmentation method as defined in any one of claims 1 to 7.

Description

Single-wood point cloud instance segmentation method, device, equipment and storage medium Technical Field The invention relates to the technical field of urban landscaping maintenance, in particular to a single-wood point cloud instance segmentation method, device and equipment and a storage medium. Background In the processes of digital landscaping maintenance, urban landscaping asset census and refined tree management, three-dimensional point cloud data acquired based on laser radar or photogrammetry is generally required to realize single-tree instance segmentation of each tree in a scene so as to further calculate key parameters such as breast Diameter (DBH), tree height, crown width, crown volume, trunk inclination angle, health state and the like, and support business applications such as greening asset statistics, pest management, growth monitoring and the like. In the prior art, the processing of the three-dimensional point cloud for the garden scene generally comprises the steps of point cloud preprocessing, target extraction, clustering or region growth and other example division steps. However, landscaping scene point clouds often have the conditions of obvious relief of topography and overlapping of tree space adjacent and crowns, so that single-wood instance segmentation is difficult to stably realize automation and high precision in engineering application, and the following technical difficulties are mainly reflected: First, topography fluctuations and ground irregularities cause trunk height references to be difficult to unify. When trunk interception is performed at different ground height positions by adopting uniform height thresholds, the inconsistent trunk interception height is easy to occur, so that the positioning of the subsequent instance initialization (such as seed point generation or seed region generation based on trunk cross section) is unstable, and the instance segmentation result drift or consistency drop is further caused. Second, the spatial proximity of trees and overlapping crowns makes single-wood boundaries difficult to determine. In the crown overlapping or branch-leaf communicating region, the problem that multiple trees are mixed or single trees are split easily occurs only by depending on an instance division strategy of local distance or density difference, and stable single-tree instance point cloud which can be used for subsequent parameter calculation is difficult to obtain. Therefore, in the prior art, there is a need for a Shan Mudian cloud instance segmentation method capable of eliminating the influence of topography fluctuation on trunk height references in a complex garden scene point cloud and realizing single-wood instance aggregation segmentation based on stable trunk seeds so as to provide reliable input data for subsequent tree parameter calculation and greening asset statistics. Disclosure of Invention The application aims to provide a single wood point cloud instance segmentation method, device, equipment and storage medium, which are used for obtaining tree point clouds and ground point clouds by carrying out semantic segmentation on urban garden scene point clouds, carrying out Z-axis standardization on the tree point clouds by taking the ground point clouds as references, further intercepting trunk height section point clouds, carrying out three-dimensional to two-dimensional generation thermodynamic diagrams, carrying out circle detection on the obtained trunk sections based on the thermodynamic diagrams, mapping the trunk sections back to three-dimensional generation root seed points, and then driving three-dimensional growth type aggregation segmentation by the root seed points to output single wood instance results, thereby reducing unstable segmentation caused by complex terrains and tree adhesion, and improving the robustness and the feasibility of single wood instance segmentation. In a first aspect, the present application provides a method for dividing a single-wood point cloud instance, including the following steps: a scene semantic segmentation step, namely executing semantic segmentation on the three-dimensional point cloud of the landscaping scene to obtain a tree point cloud and a ground point cloud; a Z-axis standardization step of estimating local ground height by taking the ground point cloud as a reference, and executing Z-axis standardization on the tree point cloud to uniformly map the ground heights at different positions to 0; The trunk height section intercepting and three-dimensional converting step, namely intercepting trunk height section point cloud in the standardized tree point cloud, and generating a two-dimensional thermodynamic diagram from three-dimensional projection of the trunk height section point cloud; Performing binarization and circle detection on the two-dimensional thermodynamic diagram to obtain trunk section circle center coordinates and radius information, and mapping the circle information back t