CN-121999180-A - Point cloud data reconstruction method in machine vision three-dimensional measurement

CN121999180ACN 121999180 ACN121999180 ACN 121999180ACN-121999180-A

Abstract

The invention relates to the technical field of machine vision three-dimensional measurement, in particular to a point cloud data reconstruction method in machine vision three-dimensional measurement, which comprises a data acquisition module, a surface generation module, a risk calculation module and a self-adaptive output module, wherein a system generates a candidate three-dimensional surface by acquiring an original sparse point cloud and inputting the original sparse point cloud into a generation network, the core of the system is that a geometrical illusion risk index representing physical inconsistency is calculated based on multi-view geometrical consistency constraint calculation, when the index is greater than or equal to a preset safety threshold, the system self-adaptively starts degradation protection, eliminates a false area and only outputs safe discrete point cloud, and the dynamic balance between data space connectivity and physical reality is realized.

Inventors

LI GUIDONG
LI DAWEI
XU JIANHUA
ZHANG TAO
ZHOU ENZE

Assignees

南京耘瞳科技有限公司

Dates

Publication Date: 20260508
Application Date: 20260331

Claims (8)

1. A method for reconstructing point cloud data in machine vision three-dimensional measurement, the method comprising: the machine vision system acquires original sparse point cloud data aiming at a target scene; The machine vision system inputs the original sparse point cloud data into a preset implicit representation-based generated neural network model, and a candidate three-dimensional surface containing speculative geometric information is generated; The machine vision system calculates the reprojection error of the candidate three-dimensional surface based on multi-view geometric consistency constraint, and calculates a geometric illusion risk index representing physical inconsistency based on the reprojection error; And the machine vision system executes self-adaptive reconstruction data output operation according to the magnitude relation between the geometric illusion risk index and a preset safety threshold.
2. The method for reconstructing point cloud data in machine vision three-dimensional measurement according to claim 1, wherein said performing an adaptive reconstruction data output operation according to a magnitude relation between the geometric illusion risk index and a preset safety threshold comprises: if the geometric illusion risk index is less than the safety threshold, the machine vision system determines that the candidate three-dimensional surface is an effective reconstruction result; and the machine vision system performs gridding processing on the candidate three-dimensional surfaces and outputs continuous three-dimensional model data.
3. The method for reconstructing point cloud data in machine vision three-dimensional measurement according to claim 1, wherein said performing an adaptive reconstruction data output operation according to a magnitude relation between the geometric illusion risk index and a preset safety threshold comprises: if the geometric illusion risk index is greater than or equal to the safety threshold, the machine vision system determines that the candidate three-dimensional surface has a generation artifact; The machine vision system starts a degradation protection strategy, and eliminates areas, which lack projection correspondence with the original sparse point cloud data, in the candidate three-dimensional surfaces; And the machine vision system outputs discrete point cloud segment data subjected to outlier noise filtering.
4. The method for reconstructing point cloud data in machine vision three-dimensional measurement according to claim 1, wherein the performing physical consistency check on the candidate three-dimensional surfaces based on multi-view geometric constraint logic, and calculating to obtain a geometric illusion risk index, comprises: The machine vision system builds a light projection path aiming at the original sparse point cloud data; The machine vision system calculates a reprojection error value of the candidate three-dimensional surface under the light projection path; And the machine vision system combines the reprojection error value and a preset rigid body motion constraint condition to generate the geometric illusion risk index.
5. The method for reconstructing point cloud data in machine vision three-dimensional measurement according to claim 4, wherein said combining said re-projection error value and a preset rigid motion constraint condition to generate said geometric illusion risk index comprises: If the reprojection error value is greater than a preset error margin, the machine vision system identifies a structured artifact region in the candidate three-dimensional surface; the machine vision system calculates geometrical structure dispersion increment of the structured artifact area on time sequence as geometrical entropy increment; and the machine vision system calculates a weighted sum of the reprojection error value and the geometric entropy increment to obtain the geometric illusion risk index.
6. A method for reconstructing point cloud data in machine vision three-dimensional measurement according to claim 3, wherein said outputting discrete point cloud segment data filtered by outlier noise comprises: The machine vision system utilizes high-confidence characteristic points in the original sparse point cloud data to construct a local rigid support structure; the machine vision system takes the local rigid support structure as an output result and marks blank state information of a non-connected region.
7. The method for reconstructing point cloud data in machine vision three-dimensional measurement of claim 1, further comprising: the machine vision system performs medium interference analysis on the original sparse point cloud data; If the original sparse point cloud data are monitored to have non-Gaussian distributed structured noise clusters, the machine vision system increases the calculation weight of the heavy projection error value when calculating the geometric illusion risk index; If the original sparse point cloud data is monitored to accord with the preset random noise distribution, the machine vision system maintains the current calculation weight.
8. The method for reconstructing point cloud data in machine vision three-dimensional measurement of claim 1, further comprising: After the machine vision system outputs the reconstruction data, acquiring observation feedback data at a subsequent moment; The machine vision system updates network parameters of the generated implicit field model online based on the deviation between the observed feedback data and the reconstructed data.

Description

Point cloud data reconstruction method in machine vision three-dimensional measurement Technical Field The invention relates to the technical field of machine vision three-dimensional measurement, in particular to a point cloud data reconstruction method in machine vision three-dimensional measurement. Background In the current machine vision three-dimensional measurement application, particularly in unstructured complex environments such as high dust, severe shielding and the like, a sensor can only acquire original sparse point cloud data with a large range of target scenes, and meanwhile, interference media such as heat haze, high-concentration chemical smoke and the like in the environment are extremely easy to generate highly complex environmental noise. In order to carry out complement reconstruction on the sparse data, the prior scheme usually introduces a generated neural network model based on implicit expression to carry out geometric information inference, and although the scheme can break through resolution limitation and fill a missing area, false structures and artifacts without actual physical support are often generated due to the fact that the model is easy to excessively infer under the interference of complex environments, and the physical reality of the presumed surface cannot be accurately estimated only by means of the confidence of an implicit field model, so that the output reconstructed data has serious physical distortion risks, and serious errors of subsequent physical interaction tasks are extremely easy to cause. Therefore, how to effectively suppress the geometric illusion of the generated model under the complex interference environment, and ensure the physical reality of the three-dimensional reconstruction data space while ensuring the connectivity of the three-dimensional reconstruction data space becomes a technical problem to be solved. Disclosure of Invention The invention aims to provide a point cloud data reconstruction method in machine vision three-dimensional measurement, which solves the technical problems that in the prior art, under the condition of facing complex data interference, a generated model is easy to excessively infer, a false structure and an artifact without actual physical support are generated, and the physical reality of a presumed surface cannot be accurately estimated only by means of the confidence of an implicit field model, so that serious physical distortion risks and the like are obviously insufficient. The aim of the invention can be achieved by the following technical scheme: the machine vision system acquires original sparse point cloud data aiming at a target scene; The machine vision system inputs the original sparse point cloud data into a preset implicit representation-based generated neural network model, and a candidate three-dimensional surface containing speculative geometric information is generated; The machine vision system calculates the reprojection error of the candidate three-dimensional surface based on multi-view geometric consistency constraint, and calculates a geometric illusion risk index representing physical inconsistency based on the reprojection error; And the machine vision system executes self-adaptive reconstruction data output operation according to the magnitude relation between the geometric illusion risk index and a preset safety threshold. Further, the performing an adaptive reconstruction data output operation according to the magnitude relation between the geometric illusion risk index and a preset safety threshold value includes: if the geometric illusion risk index is less than the safety threshold, the machine vision system determines that the candidate three-dimensional surface is an effective reconstruction result; and the machine vision system performs gridding processing on the candidate three-dimensional surfaces and outputs continuous three-dimensional model data. Further, the performing an adaptive reconstruction data output operation according to the magnitude relation between the geometric illusion risk index and a preset safety threshold value includes: if the geometric illusion risk index is greater than or equal to the safety threshold, the machine vision system determines that the candidate three-dimensional surface has a generation artifact; The machine vision system starts a degradation protection strategy, and eliminates areas, which lack projection correspondence with the original sparse point cloud data, in the candidate three-dimensional surfaces; And the machine vision system outputs discrete point cloud segment data subjected to outlier noise filtering. Further, the step of performing physical consistency check on the candidate three-dimensional surfaces based on the multi-view geometric constraint logic to obtain a geometric illusion risk index through calculation includes: The machine vision system builds a light projection path aiming at the original sparse point cloud data; T