CN-122023650-A - Differential point cloud simplifying method and system based on sensitivity guidance

CN122023650ACN 122023650 ACN122023650 ACN 122023650ACN-122023650-A

Abstract

The invention relates to a differentiating point cloud simplifying method and a differentiating point cloud simplifying system based on sensitivity guidance, wherein the method comprises the steps of obtaining an original point cloud data set; the method comprises the steps of analyzing curvature and roughness characteristics of a local neighborhood of each point cloud data, carrying out weighted fusion to obtain a sensitivity value of the point, constructing a differential point cloud simplifying network model based on sensitivity guidance, optimizing parameters of the differential point cloud simplifying network model based on sensitivity guidance, inputting the point cloud data to be simplified into the point cloud simplifying network model to output the simplified point cloud data, and evaluating performance of the simplified point cloud data. According to the invention, by introducing a sensitivity guide mechanism and a differentiable sampling network, the point cloud can be simplified in proportion flexibly, the complex geometric details are effectively reserved and model holes are avoided while the data volume is obviously reduced, and the adaptability and reconstruction quality of the point cloud processing are improved.

Inventors

LI JIAN
HU QINGWU
CUI HAO
ZHAO PENGCHENG
LI JIAYUAN
TU ZHENFA

Assignees

武汉微景易绘科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260123

Claims (10)

1. A sensitivity guidance-based differentiable point cloud reduction method, comprising: s1, acquiring an original point cloud data set, and preprocessing the original point cloud data set; S2, obtaining a sensitivity value of each point by analyzing the curvature and roughness characteristics of the local neighborhood of each point cloud data and carrying out weighted fusion; s3, constructing a sensitivity-guided differential point cloud reduced network model, wherein the sensitivity-guided differential point cloud reduced network model comprises a sensitivity weighting prediction module, a sensitivity-guided differential sampling module and a geometric reconstruction constraint module; s4, optimizing parameters of a differentiable point cloud reduced network model based on sensitivity guidance by adopting a sensitivity loss function, a repulsion loss function and a reconstruction loss function; S5, inputting the point cloud data to be reduced to a differential point cloud reduced network model based on sensitivity guidance, outputting reduced point cloud data, and evaluating the performance of the reduced point cloud data.
2. The sensitivity guidance-based differentiable point cloud reduction method of claim 1, wherein the acquiring and preprocessing the raw point cloud data comprises: Step S11, selecting 27 high-resolution models as an original point cloud data set based on Visionair libraries. Wherein the single point cloud scale of the 27 high resolution models is about Counting the number of points; Step S12, introducing noise injection and density resampling to carry out data enhancement on the original point cloud data set; step S13, a Patch division strategy is adopted for samples in an original point cloud data set; And S14, dividing the original point cloud data set according to a homologous division principle.
3. The sensitivity guidance-based differentiable point cloud reduction method of claim 1, wherein the obtaining the sensitivity value of each point by analyzing the curvature and roughness characteristics of the local neighborhood of the point cloud data and performing weighted fusion comprises: Step S21, constructing points in the original point cloud And calculating a centroid and covariance matrix of the neighborhood; Step S22, according to the point Centroid and covariance matrix of a neighborhood of (c) and computing points in an original point cloud Curvature and roughness of (a); step S23, dotting The curvature and roughness of the model are normalized and then weighted and fused to obtain points Is a sensitivity value of (a).
4. The sensitivity guidance-based differentiable point cloud reduction method of claim 3, wherein the constructing points in the original point cloud And computing the centroid and covariance matrix of the neighborhood comprises: for points By selecting a suitable neighborhood radius epsilon, the point is constructed Is a neighborhood of (a) And to the neighborhood PCA (principal component analysis) is performed if the neighborhood If the number of points contained in the neighbor is less than 3, skipping the calculation of the point, and calculating the neighbor Centroid and covariance matrix of (a) : The neighborhood is provided with Centroid of (2) The calculation formula of (2) is as follows: Wherein, the Is a neighborhood Is used to determine the centroid of the (c), Is taken as a point Is used in the neighborhood of (a), Is a neighborhood Points within; The neighborhood is provided with Covariance matrix of (2) The calculation formula is as follows: Wherein, the Is a neighborhood Is used for the co-variance matrix of (a), Is taken as a point Is used in the neighborhood of (a), Is a neighborhood The point in the inner side of the frame, Is a neighborhood Is a centroid of (c).
5. The sensitivity guidance-based differentiable point cloud reduction method of claim 1, wherein the points are reduced The curvature and roughness of the model are normalized and then weighted and fused to obtain points Comprises: the min-max normalization method is adopted to normalize the points Mapping the curvature and roughness of (1) to a range of [0,1 ]; Point(s) Curvature value of (2) The normalization formula is: Wherein, the Is taken as a point The curvature value after the normalization is used for the correction, Is taken as a point Is used for the curvature value of the lens, At the maximum value of the curvature of the web, Is the minimum value of curvature; Point(s) Roughness of (2) The normalization formula is: Wherein, the Is taken as a point The roughness after the normalization is carried out, Is taken as a point Is used for the polishing of the surface of the substrate, At the maximum value of the roughness of the surface, Is the minimum of roughness; for points Dots (dot) Sensitivity of (2) The calculation formula is as follows: Wherein, the Is taken as a point Is a sensitivity of (2); Is taken as a point The roughness after the normalization is carried out, Is taken as a point The curvature value after the normalization is used for the correction, For the weight coefficient corresponding to the roughness, For the weight coefficient corresponding to the curvature, And (2) and 。
6. The sensitivity guidance-based differentiable point cloud reduction method of claim 1, wherein the constructing the sensitivity guidance-based differentiable point cloud reduction network model comprises: step S31, a sensitivity weighted prediction module is used as a perception front end of a network, the sensitivity weighted prediction module multiplexes an improved PCPNet architecture, extracts multi-scale geometric features from an original point cloud, predicts the sensitivity of each point, outputs fusion features, and provides priority guidance for subsequent sampling; step S32, a sensitivity-guided differential sampling module guides sampling points to gather towards a high-feature area by taking sensitivity weights in fusion features as guidance, so as to obtain a simplified point cloud Q and ensure that high-sensitivity points are reserved preferentially; in step S33, the geometric reconstruction constraint module constrains the reduced point cloud Q to contain enough geometric information to recover the original shape by reconstructing the encoder, so as to force the sampling points to be distributed at geometric key positions, and ensure the integrity and representativeness of the reduced point cloud.
7. The sensitivity guidance-based differentiable point cloud reduction method of claim 1, wherein the integrated loss function is: Wherein, the In order to integrate the loss function(s), As a function of the sensitivity loss, In order to provide a repulsive force loss function, In order to reconstruct the loss function, As a weight of the sensitivity loss function, For the weight of the repulsive force loss function, Weights for reconstructing the loss function; The sensitivity loss function is: Wherein, the N is the total number of points of the original input point cloud, i is the points in the original input point cloud, As the prediction sensitivity value of the i-th point of the output, A sensitivity truth value for the ith point; The repulsive force loss function is: Wherein, the For the repulsive force loss function, Q is the target reduced point cloud point set, i is the sampling point in the reduced point cloud set Q, j is the point in the neighborhood corresponding to the point i, k is the neighborhood radius, As a penalty function for the distance correspondence, Is a three-dimensional coordinate vector of the point i, Is a three-dimensional coordinate vector of the point j, The sensitivity weight corresponding to the point i; the reconstruction loss function is: Wherein, the In order to reconstruct the loss function, For the original set of input point cloud points, For the reconstruction point cloud point set output by the decoder, x is a point in the original point cloud point set P, and y is the reconstruction point cloud point set Is a point of (2).
8. A sensitivity guidance-based differentiable point cloud reduction system, applied to the sensitivity guidance-based differentiable point cloud reduction method of claims 1-7, the system comprising: the data acquisition module is used for acquiring an original point cloud data set and preprocessing the original point cloud data set; The sensitivity calculation module is used for obtaining the sensitivity value of each point by analyzing the curvature and roughness characteristics of the local neighborhood of each point cloud data and carrying out weighted fusion; The modeling module is used for constructing a sensitivity-guiding-based differential point cloud reduced network model, and the sensitivity-guiding-based differential point cloud reduced network model comprises a sensitivity weighting prediction module, a sensitivity-guiding differential sampling module and a geometric reconstruction constraint module; the model training optimization module is used for optimizing parameters of the differential point cloud reduced network model based on sensitivity guidance by adopting a sensitivity loss function, a repulsion loss function and a reconstruction loss function; The prediction output module is used for inputting the point cloud data to be reduced to the differential point cloud reduced network model based on sensitivity guidance, outputting the reduced point cloud data, and evaluating the performance of the reduced point cloud data.
9. A network side server, comprising: at least one processor, and A memory communicatively coupled to the at least one processor, wherein, The memory stores instructions executable by the at least one processor to enable the at least one processor to perform the sensitivity guidance-based differentiable point cloud reduction method of any of claims 1 to 7.
10. A computer readable storage medium storing a computer program, wherein the computer program when executed by a processor implements the sensitivity guidance-based differentiable point cloud reduction method of any of claims 1 to 7.

Description

Differential point cloud simplifying method and system based on sensitivity guidance Technical Field The invention relates to the technical field of point cloud data processing, in particular to a differential point cloud simplifying method and system based on sensitivity guidance. Background The point cloud sampling technology is a key preprocessing step in three-dimensional computer vision and is widely applied to the fields of automatic driving, robot navigation, cultural relic protection, three-dimensional modeling and the like. The point cloud data are collected through three-dimensional sensors such as a laser radar and the like, so that the geometric structure and the surface morphology of the object can be accurately expressed. However, the original point cloud data often includes a large number of redundant points, resulting in heavy storage burden and low calculation efficiency, and thus effective point cloud reduction processing is required. Traditional point cloud reduction methods include furthest point sampling, random sampling, grid sampling and the like. Although the method can reduce the data scale, geometric details are difficult to effectively keep in the simplifying process, and particularly, the problems of feature loss, cavities and the like are easy to cause on a model with a complex structure. The prior point cloud sampling based on deep learning mainly comprises a sampling network based on scoring and a generating type sampling network, wherein the sampling network is used for screening the importance of points after scoring, and the generating type sampling network is used for directly generating the simplified point cloud. However, the existing deep learning simplifying method still has obvious defects that firstly, the reserving capacity of high-frequency detail features (such as corner points and edges) is limited, secondly, fixed sampling points are needed to be preset, adaptability to point clouds of different scales is lacking, thirdly, excessive clustering is easy to occur in a feature area, so that point clouds of non-feature areas are sparse, even geometrical cavities occur, and thirdly, the processing capacity of complex fine structures (such as a document model) is weak, and the overall structure and the local detail are difficult to combine. Therefore, there is a need to provide new sensitivity-guided differentiable point cloud reduction methods. Disclosure of Invention Based on the above-mentioned problems existing in the prior art, an object of the embodiments of the present invention is to provide a differential point cloud simplifying method and system based on sensitivity guidance, which can flexibly and proportionally simplify the point cloud by introducing a sensitivity guiding mechanism and a differential sampling network, effectively reserve complex geometric details and avoid model cavities while remarkably reducing data volume, and improve the adaptability and reconstruction quality of point cloud processing. In order to achieve the purpose, the technical scheme adopted by the invention is that the differentiating point cloud simplifying method based on sensitivity guidance comprises the following steps: s1, acquiring an original point cloud data set, and preprocessing the original point cloud data set; S2, obtaining a sensitivity value of each point by analyzing the curvature and roughness characteristics of the local neighborhood of each point cloud data and carrying out weighted fusion; s3, constructing a sensitivity-guided differential point cloud reduced network model, wherein the sensitivity-guided differential point cloud reduced network model comprises a sensitivity weighting prediction module, a sensitivity-guided differential sampling module and a geometric reconstruction constraint module; s4, optimizing parameters of a differentiable point cloud reduced network model based on sensitivity guidance by adopting a sensitivity loss function, a repulsion loss function and a reconstruction loss function; S5, inputting the point cloud data to be reduced to a differential point cloud reduced network model based on sensitivity guidance, outputting reduced point cloud data, and evaluating the performance of the reduced point cloud data. Further, the obtaining the original point cloud data and preprocessing the original point cloud data includes: Step S11, selecting 27 high-resolution models as an original point cloud data set based on Visionair libraries. Wherein the single point cloud scale of the 27 high resolution models is about Counting the number of points; Step S12, introducing noise injection and density resampling to carry out data enhancement on the original point cloud data set; step S13, a Patch division strategy is adopted for samples in an original point cloud data set; And S14, dividing the original point cloud data set according to a homologous division principle. Further, the obtaining the sensitivity value of the point by analyzing the curv