CN-121053385-B - Laser point cloud semantic segmentation method and system based on multiple random sampling

CN121053385BCN 121053385 BCN121053385 BCN 121053385BCN-121053385-B

Abstract

The invention discloses a laser point cloud semantic segmentation method and system based on multiple random sampling, comprising the steps of selecting a point cloud feature extractor, constructing a point cloud semantic segmentation model, executing multiple non-overlapping random sampling at each downsampling stage to obtain multiple point cloud subsets with complementary spatial distribution, carrying out feature coding on each point cloud subset through the point cloud feature extractor sharing weight to obtain Gao Weidian cloud features of each point cloud subset, interpolating the Gao Weidian cloud features to the input point cloud of the current level, realizing the spatial complementation fusion of multiple observation features based on an attention mechanism, enhancing feature stability, calculating local differences of the features of different point cloud subsets, constructing an information scoring function, adaptively selecting information rich points for downsampling, and generating the point cloud of the next level. The method has the advantages of high efficiency of random sampling and space structure perception capability, remarkably improves the semantic segmentation precision of the point cloud, does not depend on a specific point cloud feature extractor, has universality and expandability, and is easy to realize.

Inventors

DAI HENGMING
ZHAO ZHIFANG
LIU CHAOHAI
WANG JINQIANG

Assignees

云南大学
云南省遥感中心

Dates

Publication Date: 20260508
Application Date: 20250819

Claims (5)

1. The laser point cloud semantic segmentation method based on multiple random sampling is characterized by comprising the following steps of: Selecting a point cloud feature extractor, constructing a semantic segmentation model, extracting hierarchical features, and performing non-overlapping random sampling for multiple times in each downsampling stage to obtain a plurality of point cloud subsets with complementary spatial distribution; step 2, carrying out feature coding on each point cloud subset through a point cloud feature extractor sharing weight to obtain Gao Weidian cloud features of each point cloud subset; step 3, propagating Gao Weidian cloud features of each point cloud subset to the input point cloud of the current level through a nearest neighbor interpolation method, and fusing multiple observation features through a attention mechanism based on a spatial relationship, wherein the method comprises the following steps: The Gao Weidian cloud features extracted from each point cloud subset are transmitted to the original point cloud of the current level through nearest neighbor interpolation, so that each point obtains feature representation under a plurality of random observations, including the central feature of the point cloud subset And assist features interpolated from other point cloud subsets ; Constructing, for each point, a relative encoding based on spatial location and feature information The formula is as follows: Wherein, the Representing the coordinates of the central feature point, Representing the coordinates of the auxiliary feature points of the s-th random sampling, and the method is characterized in that ) Representing the relative positional difference of the center feature point and the auxiliary feature point, Represents the auxiliary characteristic of the s-th random sampling, and the method is that ) Representing central features And assist features Is characterized by differences in features; Inputting the relative code into a multi-layer perceptron network for generating a representation of the feature And a multi-layer perceptron network for generating attention weights ; The central feature and the auxiliary feature are fused through an attention mechanism to obtain a fused feature The calculation formula is as follows: Wherein, the Representing a multi-layer perceptron network for fine tuning, The softmax function is represented by a graph, Representing the number of random samplings, Representing element-by-element multiplication; And 4, adaptively selecting information rich points to downsample according to local difference calculation information scores of Gao Weidian cloud features of each point cloud subset to generate a next-level point cloud, wherein the local difference calculation information scores according to Gao Weidian cloud features of each point cloud subset comprise the following substeps: Constructing a high-dimensional feature set for each point, wherein the high-dimensional feature set comprises the central feature of the point cloud subset and a plurality of auxiliary observation features from adjacent random point cloud subsets, performing maximum pooling operation on each group of auxiliary observation features to obtain an aggregate feature without the central feature, and recording the aggregate feature as : Wherein, the Representing a maximum pooling operation, m represents the total number of the set of auxiliary observation features, Representing the auxiliary observation features of the group; Performing a maximum pooling operation on each set of auxiliary observation features and the central features to obtain an aggregate feature containing the central features : The following two feature distances are calculated respectively, and local differences of different point cloud subset features are obtained: Wherein, the Representation for reflecting central characteristics A value of importance of (a); Representation for reflecting central characteristics A value of the degree of influence on the maximum pooled output; Based on And (3) with An information scoring function for measuring the difference of the local features is constructed, and the information scoring function is formed as follows: Wherein, the Information scores reflecting the local structural complexity and information-rich conditions of the point cloud points are represented.
2. The method for semantic segmentation of laser point cloud based on multiple random samplings of claim 1, wherein said point cloud feature extractor is selected according to practical conditions including but not limited to PointNet ++, randLA-Net, KPConv.
3. The method for semantic segmentation of laser point clouds based on multiple random sampling according to claim 1, wherein the random sampling is not a back sampling, the obtained point cloud subsets are not overlapped, all the point cloud subsets are complementary in spatial distribution, and the input point clouds of the current level can be completely covered.
4. The method for semantic segmentation of laser point cloud based on multiple random samplings according to claim 1, wherein said adaptively selecting information rich points for downsampling comprises: And taking the information score as a weight, and executing polynomial sampling for finishing the point cloud downsampling operation of the information retention and the region balance.
5. A laser point cloud semantic segmentation system based on multiple random sampling, which uses the laser point cloud semantic segmentation method based on multiple random sampling as set forth in any one of claims 1-4, and is characterized by comprising a multiple random sampling module, a feature extraction module, a multiple observation feature fusion module and an adaptive downsampling module: The multi-time random sampling module is used for performing multi-time non-overlapping random sampling on the input point clouds of each level in the hierarchical feature extraction to generate a plurality of point cloud subsets with complementary spatial distribution; The feature extraction module is used for carrying out feature coding on the point cloud feature extractor of the shared weight for each point cloud subset to obtain Gao Weidian cloud features of each point cloud subset; The multi-observation feature fusion module is used for transmitting Gao Weidian cloud features of each point cloud subset to the input point cloud of the current level through a nearest neighbor interpolation method and fusing the multi-observation features through a spatial relationship-based attention mechanism; the self-adaptive downsampling module is used for calculating information scores according to local differences of Gao Weidian cloud features of each point cloud subset, adaptively selecting information rich points to downsample, and generating a next-level point cloud.

Description

Laser point cloud semantic segmentation method and system based on multiple random sampling Technical Field The invention relates to the technical field of laser point cloud semantic segmentation, in particular to a laser point cloud semantic segmentation method and system based on multiple random sampling. Background Semantic segmentation of the original laser point cloud is a key premise of follow-up tasks such as city modeling, digital elevation model construction, natural resource investigation and the like. In the end-to-end point cloud processing method, random sampling is a widely used strategy for constructing a hierarchical structure of point clouds in the feature extraction process. This has the advantage of efficient computation, easy extension to large-scale data sets. However, random sampling also has the obvious problems that the same input can generate characteristic representations with obvious difference under different sampling due to randomness in the sampling process, the optimization process and the robustness of a model are affected, and key region characteristics are lost or an oversampled dense region is easily caused without considering the distribution rule of point clouds in space, so that the overall segmentation performance is affected. Therefore, a point cloud processing method for improving the stability and spatial representation capability of random sampling features without significantly increasing the computational complexity is needed. Disclosure of Invention The invention aims to solve the problems of unstable characteristics, lack of spatial structure perception in sampling and the like of a laser point cloud semantic segmentation method in the prior art when a random sampling strategy is adopted, and provides a laser point cloud semantic segmentation method and a laser point cloud semantic segmentation system based on multiple random sampling. In order to solve the technical problems, the invention provides the following technical scheme: a laser point cloud semantic segmentation method based on multiple random sampling comprises the following steps: Selecting a point cloud feature extractor, constructing a semantic segmentation model, extracting hierarchical features, and performing non-overlapping random sampling for multiple times in each downsampling stage to obtain a plurality of point cloud subsets with complementary spatial distribution; step 2, carrying out feature coding on each point cloud subset through a point cloud feature extractor sharing weight to obtain Gao Weidian cloud features of each point cloud subset; step 3, propagating Gao Weidian cloud features of each point cloud subset to the input point cloud of the current level through a nearest neighbor interpolation method, and fusing multiple observation features through a attention mechanism based on a spatial relationship so as to improve stability and consistency of feature extraction; and 4, calculating information scores according to local differences of Gao Weidian cloud features of each point cloud subset, adaptively selecting information rich points for downsampling, and generating a next-level point cloud. As a preferred scheme of the laser point cloud semantic segmentation method and system based on multiple random sampling, the point cloud feature extractor can be selected according to practical conditions, including but not limited to PointNet ++, randLA-Net and KPConv. As the optimal scheme of the laser point cloud semantic segmentation method and the system based on multiple random sampling, the random sampling is not to put back the sampling, the obtained point cloud subsets are not overlapped, all the point cloud subsets are complementary in spatial distribution, and the input point cloud of the current level can be completely covered. As a preferred scheme of the laser point cloud semantic segmentation method and system based on multiple random sampling, the method for propagating Gao Weidian cloud features of each point cloud subset to the input point cloud of the current level through a nearest neighbor interpolation method and fusing multiple observation features through a attention mechanism based on a spatial relationship comprises the following substeps: The Gao Weidian cloud features extracted from each point cloud subset are transmitted to the original point cloud of the current level through nearest neighbor interpolation, so that each point obtains feature representation under a plurality of random observations, wherein the feature representation comprises a central feature f c of the point cloud subset and auxiliary features f o obtained through interpolation of other point cloud subsets; For each point, a relative code F r based on spatial location and feature information is constructed, as follows: wherein x c denotes the coordinates of the central feature point, Representing the coordinates of the auxiliary feature point of the s-th random sampling,Representing the relati