Search

CN-121999274-A - Urban vehicle-mounted point cloud forest classification method based on RoPE operator local feature aggregation

CN121999274ACN 121999274 ACN121999274 ACN 121999274ACN-121999274-A

Abstract

The invention relates to a method for classifying urban vehicle-mounted point cloud forest based on RoPE operator local feature aggregation, belonging to the technical field of photogrammetry and remote sensing. The invention adopts the ROPE operator technology to encode the relative coordinates of each point in the neighborhood and the original characteristics to establish a weight relation, helps a model to distinguish trunk fine characteristics, utilizes the encoding operator to dynamically aggregate secondary local characteristics, captures the local structure and the context information of the tree, solves the problem of low classification precision of the trunk and the blocked tree, introduces a U-shaped encoder-decoder network architecture and a jump connection strategy, fuses the tree characteristics among different levels, and realizes the acquisition of multi-scale tree information. Compared with other methods, the method can solve the problems that trunk classification errors, obstacle classification errors and the like often occur in urban forest classification, and achieve accurate segmentation of complete forest information in complex urban scenes.

Inventors

  • ZHAO QIAN
  • WU XINMING
  • LIU XUDONG
  • LONG ZHAOXIN
  • ZOU KANG
  • WANG XUN
  • JIANG ZEWEI
  • FAN HONGYING
  • MENG QINGAN

Assignees

  • 西南技术物理研究所

Dates

Publication Date
20260508
Application Date
20251225

Claims (10)

  1. 1. A method for classifying urban vehicle-mounted point cloud forest based on RoPE operator local feature aggregation is characterized by comprising the following steps of: Step 1, three-dimensional laser point cloud data of a forest are obtained, noise reduction processing and sample labeling are carried out on the three-dimensional laser point cloud data, and finally point clouds in the three-dimensional laser point cloud data are connected with a full connection layer to learn the spatial characteristics of each point cloud; step 2, performing relative position coding on the space features by using RoPE operators at each coding layer, and endowing the space features with relative coordinates and unique weights of mapping features; Step 3, forming a multi-scale local feature aggregation module by utilizing two RoPE operators, and dynamically capturing the local geometric features of the point clouds, wherein in the step, the operation of step 2 is executed again to fuse the fused features again, the number of the input point clouds is gradually reduced after the point clouds are processed by each coding layer, only one fourth of the original number is reserved, and simultaneously, the features of each point cloud are mapped into a feature space with higher dimension; And 4, fusing the high-level features with the detail features from the shallow layers through cross-layer connection of each decoding layer, and finally realizing semantic segmentation of the forest to obtain a classification result.
  2. 2. The method according to claim 1, wherein step 1 is specifically as follows: step 1.1, acquiring three-dimensional laser point cloud data and storing the data in the form of three-dimensional coordinates (x, y, z) of points; step 1.2, deleting isolated points with fewer points in a local area; Step 1.3, marking the data by utilizing a point cloud visualization tool, and distributing corresponding labels; step 1.4, performing decentration treatment on the three-dimensional laser point cloud data; Step 1.5, mapping the original three-dimensional coordinates (x, y, z) to an 8-dimensional spatial feature by using the full connection layer.
  3. 3. The method according to claim 1, wherein step 1.4 is specifically as follows: for a given point cloud First calculate its centroid The calculation formula is as follows: c= For each point cloud Using the formula Obtaining the point cloud after the decentralization treatment N is the number of point clouds.
  4. 4. The method according to claim 1, wherein step 2 is specifically as follows: step 2.1, searching n neighborhood points of each center point by using a nearest neighbor search algorithm; step 2.2, obtaining the characteristic information of the relative distance and the relative characteristic of the neighborhood point relative to the center point; step 2.3, using periodic functions Absolute position coding is performed on the even channel characteristics of the input layer, pos represents the current position, i represents the current dimension, Representing the dimensions of the model; Step 2.4, using periodic functions Absolute position coding is carried out on the odd number channel characteristics of the input layer; step 2.5, coding the position information and the characteristic information of the center point and the neighborhood point by using RoPE operators, and mapping discrete information into a uniform periodic space; Step 2.6, fusing the feature coded by RoPE operators with the original spatial feature of the point cloud; and 2.7, processing the fusion characteristics through a multi-layer perception mechanism.
  5. 5. The method of claim 4, wherein in step 2.5, the spatial features of geometrically encoded input points and their neighboring points are RoPE encoded, and a weight relationship between the features is established.
  6. 6. The method of claim 4, wherein in step 2.5, an angle parameter is calculated using RoPE operators The angle parameter is changed along with the position change of the point cloud, and the calculated angle parameter is used And finally multiplying the characteristic vector of each point cloud with the rotation matrix to enable the characteristic vector of each point cloud to rotate in the characteristic space, thereby realizing the encoding of the relative position information of each point cloud.
  7. 7. The method of claim 1, wherein in step 4, the cross-layer connection is to match and splice a feature map extracted by downsampling in an encoding stage and a feature map generated by upsampling in a decoding stage in a channel dimension.
  8. 8. The method of claim 1, wherein step 4 gradually increases the number of input point clouds and gradually decreases the dimensions during the cross-layer connection.
  9. 9. The method of claim 8, wherein in step 4, after cross-layer connection, classification prediction results (N ', nclass) are output by using two full-connection layers and one regularization layer, where N' represents the number of original point clouds and nclass represents the category to which each point cloud belongs, so as to implement classification of the forest.
  10. 10. Use of a method according to any one of claims 1 to 9 in photogrammetry and remote sensing.

Description

Urban vehicle-mounted point cloud forest classification method based on RoPE operator local feature aggregation Technical Field The invention belongs to the technical field of photogrammetry and remote sensing, and particularly relates to an urban vehicle-mounted point cloud forest classification method based on RoPE operator local feature aggregation. Background The urban forest resources have important roles in improving urban ecological environment and improving resident life quality, and accurately identifying and extracting the forest resources in the cities is also one of key links of urban ecological environment fine management, greening planning, ecological monitoring and intelligent garden construction. The foundation laser point cloud has high precision and penetrability, has the function of capturing the structural characteristics of the lower layer of the forest and the lower part of the crown from the ground view angle, and is widely applied to three-dimensional semantic segmentation of the forest. The traditional method utilizes specific attributes or preset rules of the foundation point cloud data to distinguish different objects of the point cloud data, but the method has the defects of large workload, low calculation efficiency and low automation degree. With the progress of computer technology and the rising of a deep learning method, the urban forest point cloud classification technology based on deep learning gradually becomes a new trend in the field of point cloud forest segmentation due to the advantages of high precision, high efficiency, high automation degree and the like. PointNet ++ extracts the forest features by applying PointNet in each local area so that it can better understand the local relationship of the forest data. PointNet ++ is capable of capturing local features, but there is no efficient strategy to fuse and exploit different regional features, resulting in feature extraction still being difficult when faced with trunk targets with very few points. RandLA-Net extracts and fuses local spatial forest features through an MLP (full connection layer) and a local feature aggregation layer, so that tree structure information is captured. However, in practical application, the local feature extraction mechanism is not flexible enough for a forest scene with large scale difference, and a local aggregation window with a fixed size may cause the characteristics of the blocked forest to be lost or excessively aggregated, so that the understanding and generalization capability of the model to the blocked forest are affected. In summary, in the existing method, the extraction precision of the target object of the fine trunk is insufficient in the face of the base point cloud data with high resolution but limited visual field, and the capability of distinguishing and processing the blocked forest is lacking, so that the phenomena of trunk classification error, blocked forest classification error and the like often occur in the classification of the target object of the urban forest. Disclosure of Invention First, the technical problem to be solved Aiming at the problems that the existing network has insufficient trunk extraction precision and the capability of distinguishing and processing the blocked forest, the classification errors of the trunk and the blocked forest often occur in the urban forest classification, and the accurate segmentation of the complete forest information in a complex urban scene is difficult to realize, the invention provides the urban vehicle-mounted point cloud forest classification method. (II) technical scheme In order to solve the technical problems, the invention provides a method for classifying urban vehicle-mounted point cloud forest based on RoPE operator local feature aggregation, which is realized based on a attention generation network model and comprises the following steps: Step 1, three-dimensional laser point cloud data of a forest are obtained, noise reduction processing and sample labeling are carried out on the three-dimensional laser point cloud data, and finally point clouds in the three-dimensional laser point cloud data are connected with a full connection layer to learn the spatial characteristics of each point cloud; step 2, performing relative position coding on the space features by using RoPE operators at each coding layer, and endowing the space features with relative coordinates and unique weights of mapping features; Step 3, forming a multi-scale local feature aggregation module by utilizing two RoPE operators, and dynamically capturing the local geometric features of the point clouds, wherein in the step, the operation of step 2 is executed again to fuse the fused features again, the number of the input point clouds is gradually reduced after the point clouds are processed by each coding layer, only one fourth of the original number is reserved, and simultaneously, the features of each point cloud are mapped into a featur