3D Point Cloud Label Learning Using Supervoxel Clustering

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing methods for estimating labels in three-dimensional point clouds are limited by the number of points they can handle and struggle with assigning labels to natural objects that do not form sweeping shapes, such as tree leaves.

Innovation Solution

A system that includes a ground-based height calculation unit, an intensity-RGB conversion unit, a supervoxel clustering unit, and a deep neural network learning unit, which processes three-dimensional point clouds with ground-based heights and attribute information to cluster points into supervoxels and learn labels, enabling the handling of large-scale point clouds and labeling of various object types.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If the number of points in the input point cloud is limited, then the processing speed and computational load are controlled, but the ability to handle large-scale point clouds with unlimited range is reduced

Engineering Contradiction:

Improveprocessing speedVSAvoidability to handle large-scale point clouds

Core Design Contradiction:

ProductivityVSAdaptability or versatility

Solution Approach 1:

The point cloud is divided into multiple local regions or patches, which are processed independently and then aggregated. This segmentation allows the system to handle large-scale point clouds by breaking them into manageable segments while maintaining the ability to process unlimited ranges through hierarchical aggregation of segment results.

Inventive Principle:

Principle #1Segmentation

2Device complexity

If the point cloud is processed without considering ground-based height, then the processing is simpler, but the accuracy of label estimation for objects at different elevations is reduced

Engineering Contradiction:

Improveprocessing complexityVSAvoidlabel estimation accuracy

Core Design Contradiction:

Device complexityVSMeasurement precision

Solution Approach 1:

Ground-based height calculation is performed as a preliminary step before main processing. By pre-calculating and storing ground-based height information, the system eliminates the need for complex real-time calculations during label estimation, while significantly improving accuracy for objects at different elevations through the use of this pre-computed reference data.

Inventive Principle:

Principle #10Preliminary action

3Productivity

If only position information is used for label estimation, then the processing is faster and simpler, but the accuracy of label estimation is reduced

Engineering Contradiction:

Improveprocessing efficiencyVSAvoidlabel estimation accuracy

Core Design Contradiction:

ProductivityVSMeasurement precision

Solution Approach 1:

Multiple attribute information types (position, intensity, ground-based height, and other features) are merged into a unified feature vector for label estimation. This combination allows the system to maintain processing efficiency through integrated processing while achieving superior accuracy by leveraging complementary information from multiple attributes simultaneously.

Inventive Principle:

Principle #5Merging (Combining)

4Quantity of substance

If traditional clustering methods are used, then the computational requirements are lower, but the ability to handle large-scale point clouds with complex structures is reduced

Engineering Contradiction:

Improvecomputational resourcesVSAvoidability to handle complex point cloud structures

Core Design Contradiction:

Quantity of substanceVSAdaptability or versatility

Solution Approach 1:

The clustering approach is extended from traditional two-dimensional feature spaces to multi-dimensional spaces by incorporating ground-based height, intensity, and other attributes. This dimensional expansion enables the system to handle complex point cloud structures more effectively while maintaining computational efficiency through optimized multi-dimensional indexing and search algorithms.

Inventive Principle:

Principle #17Another dimension (Dimensionality change)

Data Source

PatentUS11334744B2Three-dimensional point cloud label learning device, three-dimensional point cloud label estimating device, three-dimensional point cloud label learning method, three-dimensional point cloud label estimating method, and program

Publication Date: 2022.05.17 NIPPON TELEGRAPH & TELEPHONE CORP

AI summary

A large-scale point cloud having no limitation on the range or the number of points is set as an object, and labels are attached to the points constituting the object regardless of the type of object.A three-dimensional point cloud label learning apparatus 10A includes a ground-based height calculation unit that receives a three-dimensional point cloud and a ground surface height as inputs and outputs a three-dimensional point cloud with a ground-based height, an intensity-RGB conversion unit that receives the three-dimensional point cloud with the ground-based height as an input and outputs an intensity-RGB converted three-dimensional point cloud with the ground-based height, a supervoxel clustering unit that receives the intensity-RGB converted three-dimensional point cloud with the ground-based height, a point cloud label for learning, and a clustering hyperparameter as inputs and outputs supervoxel data with a correct answer label, and a deep neural network learning unit that receives the supervoxel data with the correct answer label and a deep neural network hyperparameter as inputs and outputs a learned deep neural network parameter.