CN-121999230-A - Point cloud single wood instance segmentation method and system based on soft clustering and Transformer

CN121999230ACN 121999230 ACN121999230 ACN 121999230ACN-121999230-A

Abstract

The invention discloses a point cloud single wood instance segmentation method and system based on soft clustering and Transformer. The method effectively solves the problem of overlarge calculation amount of large-scale forest point cloud data, adopts a soft clustering module to calculate soft attribution weight by a probabilistic attribution mechanism, strengthens the characteristic characterization capability of a boundary area, overcomes the defect that the boundary fuzzy area such as branch intersection and shielding is easy to cause example boundary distortion in the traditional hard clustering method, avoids the problems of excessive combination of examples, fuzzy boundary, high calculation cost and the like in the traditional method, remarkably improves the precision, robustness and multi-scene adaptation capability of single-wood example segmentation, and can meet the actual requirements of accurate forestry and ecological monitoring in complex forest scenes.

Inventors

XU SHENG
XIE YUHANG
XIA SHAOBO
YANG HONGXIN
SUN SHUHONG

Assignees

南京林业大学

Dates

Publication Date: 20260508
Application Date: 20260305

Claims (8)

1. A point cloud single wood instance segmentation method based on soft clustering and Transformer is characterized by comprising the following steps: S1, acquiring single-wood laser radar point cloud data, preprocessing the point cloud data and extracting hierarchical features to obtain point-by-point features; s2, calculating soft attribution weight based on the space distance and the feature semantic similarity according to the point-by-point features, and generating super-point level features; and S3, decoding and predicting the super-point level features, and outputting a single wood instance segmentation result by combining the multi-task joint loss function optimization model.
2. The method for partitioning point cloud single wood instances based on soft clustering and Transformer according to claim 1, wherein the method for preprocessing the point cloud data in S1 comprises: acquiring single wood laser radar point cloud data by adopting an airborne, vehicle-mounted or knapsack type mobile laser scanning system to obtain original point cloud data; and carrying out voxelization processing on the original point cloud data according to the original point cloud data, converting the irregular point cloud into a regular voxel grid input format, and calculating the index of the voxel to which the point belongs.
3. The method for partitioning point cloud single wood instances based on soft clustering and transformation according to claim 2, wherein in S1, the method for extracting hierarchical features comprises: Extracting point cloud layer grading characteristics by adopting sparse 3D U-Net alternately formed by sub-manifold sparse convolution and standard sparse convolution according to a regular voxel grid input format; According to the point cloud hierarchical characteristics, the receptive field is enlarged by means of layer-by-layer downsampling of an encoder, and the multiscale characteristics are fused by means of upsampling and jump connection of a decoder, so that the point-by-point characteristics are obtained.
4. The soft cluster and transform based point cloud mono instance segmentation method according to claim 1, wherein the S2 comprises: Selecting M=1000 initial super-point centers by adopting the furthest point sampling strategy according to the point-by-point characteristics, executing ball inquiry with radius of each super-point center, and gathering neighborhood points to form an initial super-point cluster; According to the initial super point cluster, calculating soft attribution weights of the points and the super points, wherein the soft attribution weights are calculated based on the similarity of the space distance and the characteristic semantics; And aggregating the super-point characteristics according to the soft attribution weight to generate super-point level characteristics.
5. The method for partitioning point cloud single-wood instances based on soft clustering and Transformer according to claim 1, wherein before decoding and predicting the super-point level features, further comprising inputting the super-point level features into a dual-branch preprocessing structure, outputting mask perception features by a lightweight MLP through mask branches, and obtaining instance branch features by linear projection through instance branches.
6. The method for partitioning point cloud single-wood instances based on soft clustering and transformation according to claim 5, wherein in S3, the method for decoding super-point level features comprises: According to the mask perception feature and the example branch feature, decoding by adopting an optimized transducer decoder, wherein the optimized transducer decoder optimizes the attention interaction sequence of a standard transducer decoder, and performs cross attention first and then self attention; According to the cross attention, a scaled dot product attention mechanism is adopted, and a soft mask matrix is introduced, wherein the soft mask matrix is generated based on a previous layer mask prediction result and is used for inhibiting background super-dot interference.
7. The soft cluster and transform based point cloud single wood instance segmentation method according to claim 6, wherein the method for predicting the point level features in S3 comprises: performing mask weighted aggregation on the super-point features through a mask perception pre-measurement head according to the decoded super-point features to obtain mask perception instance characterization; And according to the mask perceived instance characterization, completing instance classification, mask prediction and IoU quality evaluation synchronously, and outputting a single wood instance segmentation result.
8. A soft cluster and transform based point cloud unigram instance segmentation system for implementing the method of any of claims 1-7, comprising: the acquisition module is used for acquiring single-wood laser radar point cloud data, preprocessing the point cloud data and extracting hierarchical features to obtain point-by-point features; the generating module is used for calculating soft attribution weight based on the space distance and the feature semantic similarity according to the point-by-point features and generating super-point level features; And the segmentation module is used for decoding and predicting the super-point level features, combining the multi-task joint loss function optimization model and outputting a single wood instance segmentation result.

Description

Point cloud single wood instance segmentation method and system based on soft clustering and Transformer Technical Field The invention relates to the technical field of digital 3D (three-dimensional) refined modeling, in particular to a point cloud single wood instance segmentation method and system based on soft clustering and Transformer. Background The single-tree three-dimensional example segmentation is one of research hotspots in the fields of computer graphics, remote sensing science and forestry engineering, and the model reconstruction technology has been widely applied in the aspects of virtual reality simulation, ancient tree name wood digital protection, accurate forestry resource investigation, ecological environment monitoring evaluation, intelligent agriculture and forestry management and the like. In the vegetation three-dimensional modeling and structure analysis field, single wood instance segmentation is a core premise of single wood space structure analysis, is a basis for single wood three-dimensional morphology visualization and organ-level phenotype parameter extraction, and can comprehensively reflect the space topological relation, geometric morphology characteristics and distribution rule of single wood. The technology is a key element of three-dimensional point cloud semantic understanding, stand structure inversion and biomass estimation, and provides important data support for forest phenotype group study and forestry intelligent management. Compared with a two-dimensional remote sensing image, the laser radar technology has unique advantages in the aspect of acquiring vegetation three-dimensional space information, and can realize key tasks such as tree crown contour division, single wood counting statistics, fine branch and leaf separation, single wood instance division and the like. Common laser scanning systems include Airborne Laser Scanning (ALS), mobile Laser Scanning (MLS), and Terrestrial Laser Scanning (TLS). In recent years, backpack type mobile laser scanning systems and handheld laser scanning systems are mature, can collect data flexibly in complex forest environments, and provide Shan Mudian cloud information with high density and high precision. However, the point cloud data acquired by the system has the problems of non-uniformity, serious crown overlapping, complex branch topological relation and the like, and particularly the problems of noise interference and structural shielding of the point cloud data are more prominent in complex scenes such as mountain bamboo forests, canopy-and-trunk forests and the like, so that a plurality of challenges are brought to single wood instance segmentation. In single-wood instance segmentation, the core problem to be solved is how to accurately distinguish overlapping single-wood boundaries, capture fine-limb features, and avoid the problem of excessive merging or breaking of instances in the traditional method. At present, the single-wood point cloud segmentation method is mainly divided into four major categories of a traditional geometric method, a clustering method, an image fusion method and a transform-based method. The traditional geometric method realizes segmentation through strategies such as region growth and boundary detection based on prior geometric features and topological relations of trees, but the fixed parameter setting is difficult to adapt to the structure differences of branches of different tree species and different growth stages, and has limited segmentation effect on densely overlapped crowns and tiny branches in complex forest scenes. The clustering method realizes single wood clustering by excavating the similarity characteristics of the point clouds, can alleviate the over-segmentation problem caused by crown overlapping, but has insufficient characteristic capturing capability on the sparse point clouds, the clustering result is easily influenced by the density distribution of the point clouds, and the fine hierarchical segmentation of the branches is difficult to realize. The image fusion method utilizes rich semantic features of the two-dimensional image to assist in point cloud segmentation, so that the three-dimensional labeling cost is reduced, but the method is highly dependent on the registration accuracy of cross-mode data, and the registration error can directly lead to the reduction of the segmentation accuracy. The transform-based method realizes end-to-end single-wood instance segmentation by virtue of strong global context feature capturing capability, but the method has the common problems of high calculation cost and sensitivity to small sample data, and is easy to cause inaccurate instance positioning due to invalid interaction of an attention mechanism when a boundary fuzzy region is processed. To obtain high precision single-wood instance segmentation results, existing methods typically require steps in combination with post-processing operations, such as non-maxima suppressio