3D Scene Modeling with Curvature-Based Point-Cloud Segmentation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods struggle to reliably segment complex scenes with many objects and objects of complex shapes from three-dimensional point clouds, often leading to over-segmentation or under-segmentation, making it difficult to derive accurate three-dimensional models for tasks like CAD, CAE, and CAM.

Innovation Solution

A method involving curvature-based and region-growing segmentations, combined with deselection criteria, is used to segment point clouds into clusters representing objects. This includes generating parametric surfaces, determining local principal curvature, and using region-growing methods to form clusters, with deselection of less relevant parts based on curvature and spatial properties.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If curvature-based segmentation is used to segment point clouds, then segmentation can be performed based on surface characteristics, but over-segmentation occurs when many different surface curvatures are present in the scene

Engineering Contradiction:
Improvesegmentation accuracyVSAvoidnumber of segmented entities
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent applies multi-stage segmentation by first performing curvature-based segmentation to divide the point cloud into groups based on surface curvature characteristics, then performing region-growing segmentation on selected groups to form clusters, and finally performing a second region-growing segmentation on the remaining point cloud. This hierarchical segmentation approach prevents over-segmentation by combining multiple segmentation strategies.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies different segmentation strategies to different regions of the point cloud based on their local characteristics. Curvature-based segmentation is applied to identify regions with specific surface properties, while region-growing methods are applied to other regions. This allows the segmentation process to adapt to local variations in surface curvature and object geometry.

Inventive Principle:
Principle #3Local quality

2Ease of operation

If region-growing segmentation is used to segment point clouds, then objects can be clustered based on spatial continuity, but under-segmentation occurs when clusters comprise several objects or more parts than intended

Engineering Contradiction:
Improveautomatic segmentationVSAvoidobject identification accuracy
Core Design Contradiction:
Ease of operationVSMeasurement precision

Solution Approach 1:

The patent performs region-growing segmentation in multiple stages: first on selected groups obtained from curvature-based segmentation, then on the remaining point cloud. By dividing the region-growing process into separate stages applied to different subsets of the point cloud, the method prevents under-segmentation while maintaining automatic operation.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary curvature-based segmentation to identify and select specific groups of points before applying region-growing segmentation. This preliminary action prepares the data by pre-identifying regions that should be processed separately, thereby preventing under-segmentation in subsequent region-growing steps.

Inventive Principle:
Principle #10Preliminary action

3Reliability

If multiple segmentation methods are combined to improve segmentation accuracy, then object identification becomes more reliable, but the complexity of the segmentation process increases

Engineering Contradiction:
Improveobject identification reliabilityVSAvoidsegmentation process complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent divides the complex segmentation process into distinct, sequential stages: curvature-based segmentation, selection of groups, region-growing segmentation on selected groups, deselection of certain clusters, and region-growing segmentation on remaining points. This structured multi-stage approach manages complexity by breaking down the overall task into manageable steps with clear inputs and outputs at each stage.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent selectively discards certain first clusters based on predefined criteria before performing the second region-growing segmentation. This discarding step simplifies the remaining point cloud by removing already-processed or irrelevant regions, thereby reducing the complexity of subsequent segmentation operations while maintaining overall reliability.

Inventive Principle:
Principle #34Discarding and recovering

Data Source

PatentEP4625320A1Deriving a three-dimensional model of a scene
Publication Date: 2025.10.01 SIEMENS AG
  • EP4625320A1 patent drawingFigure 1
  • EP4625320A1 patent drawingFigure 2~3
  • EP4625320A1 patent drawingFigure 4

AI summary

A method for deriving a three-dimensional model of a scene (1) is provided, wherein the scene comprises at least one object (01-05). The method comprises the following steps: a) performing a physical measurement on the scene (1), which leads to a three-dimensional point cloud (PC), b) generating at least one surface (S), in particular a parametric surface, from points of the point cloud (PC) and determining a respective local principal curvature (k) for the individual points of the point cloud (PC), c) performing a first curvature-based segmentation of the point cloud (PC) into at least two groups of points (g1, g2) based on at least one threshold value (t) of the local principal curvature (k), d) performing a first region-based segmentation of at least one selected group (g1) of the groups of points (g1, g2) obtained in step c), said segmentation being based on a region-growing method and leading to a plurality of first clusters (c1a, c1b), e) determining a remaining point cloud (RPC) by deselecting a sub-set of these first clusters (c1b) based on a pre-defined deselection criterion and discarding this sub-set (c1b) from the point cloud (PC) at least for the following step f), f) determining a plurality of second clusters (c2a, c2b) by performing a second region-based segmentation of the remaining point cloud (RPC), said segmentation being based on a region-growing method. Furthermore, a corresponding computer program product for carrying out such a method is provided.