3D Point Cloud Classification via Hierarchical Semantic Fusion

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Current three-dimensional point cloud classification methods, especially those based on deep learning, face challenges in achieving high accuracy and robustness, particularly when dealing with noisy data, partial occlusions, and varying numbers of sampling points.

Innovation Solution

A full-automatic classification method for three-dimensional point clouds using a deep neural network model that integrates prior knowledge of overall point cloud categories and single semantic categories of each point. This method involves down-sampling, up-sampling, and splicing fusion of global features to improve classification accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Extent of automation

If deep learning is used for point cloud classification, then automation and intelligence are improved, but classification accuracy is reduced due to ignoring overall semantics and single point semantics

Engineering Contradiction:

Improveautomation of point cloud classificationVSAvoidclassification accuracy

Core Design Contradiction:

Extent of automationVSMeasurement precision

Solution Approach 1:

The patent segments the point cloud classification task into two distinct semantic levels: overall point cloud category semantics and single point semantic categories. By dividing the classification problem into hierarchical segments, the model can simultaneously capture global context and local details, resolving the contradiction between automation and accuracy.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent merges two types of semantic information (overall category semantics and single point semantics) into a unified classification framework. This combination allows the deep learning model to leverage both global context and local features, improving classification accuracy while maintaining automation.

Inventive Principle:

Principle #5Merging (Combining)

2Productivity

If direct recognition results are used as final prediction, then processing speed is improved, but robustness for missing point cloud conditions is reduced

Engineering Contradiction:

Improveprocessing speedVSAvoidrobustness to missing point cloud

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

The patent performs preliminary classification at the single point level before finalizing the overall category prediction. This preliminary action allows the model to build a foundation of local semantic understanding that can compensate for missing or incomplete point cloud data, improving robustness without significantly sacrificing processing speed.

Inventive Principle:

Principle #10Preliminary action

3Measurement precision

If more sampling points are used, then classification accuracy is improved, but influence of sampling point number on accuracy increases

Engineering Contradiction:

Improveclassification accuracyVSAvoidsensitivity to sampling point number

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The patent changes the approach from directly using sampling point coordinates to using down-sampled global features that capture essential characteristics regardless of the original sampling density. This parameter transformation makes the classification less sensitive to the number of sampling points while maintaining accuracy.

Inventive Principle:

Principle #35Parameter changes

4Adaptability or versatility

If point cloud data contains noise or occlusion, then real-world applicability is maintained, but recognition effect of point-based methods is reduced

Engineering Contradiction:

Improvereal-world applicabilityVSAvoidrecognition effect

Core Design Contradiction:

Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The patent uses down-sampling as a disposable preprocessing step that creates a simplified representation of the point cloud. This temporary reduction in data complexity helps filter out noise and occlusion effects, allowing the model to focus on essential features while maintaining real-world applicability.

Inventive Principle:

Principle #27Cheap short-living objects (Disposable)

Data Source

PatentUS12307741B2Full-automatic classification method for three-dimensional point cloud and deep neural network model

Publication Date: 2025.05.20 BEIJING UNIV OF CIVIL ENG & ARCHITECTURE

AI summary

A full-automatic classification method for a three-dimensional point cloud, including: acquiring a three-dimensional point cloud dataset; performing down-sampling on a three-dimensional point cloud represented by the three-dimensional point cloud dataset, selecting some points in the three-dimensional point cloud as sampling points, constructing a point cloud area group based on each sampling point, extracting a global feature of each point cloud area group, and replacing the point cloud area group where the sampling point is located with the sampling point; performing up-sampling on the three-dimensional point cloud, and performing splicing fusion on the global features of the point cloud area group where each point in the three-dimensional point cloud is located; performing category discrimination on each point in the three-dimensional point cloud; performing statistics on the number of points contained in each category, and selecting the category with the largest number of points as the category of the three-dimensional point cloud.