3D Point Cloud Component Detection via Feature Vector Matching

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current automatic inspection systems face challenges in efficiently detecting and verifying components within complex products, particularly in scenarios where large datasets are not readily available, and require extensive manual tuning and calibration, limiting their scalability and ease of maintenance.

Innovation Solution

A computer-implemented automatic inspection method that computes target and component feature vectors from 3D point-cloud representations, performs matching using a pre-trained deep neural network, and determines component presence without the need for labeling, allowing for geometric transformation and feature vector matching to verify component location and orientation within specifications.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If Deep Learning models are used for AI-based computer vision tasks, then detection accuracy and pattern recognition capability are improved, but the need for large amounts of training data and extensive manual tuning increases

Engineering Contradiction:
Improvedetection accuracyVSAvoidtraining data requirement
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent applies preliminary action by pre-training the Deep Learning model on general 3D geometric features before applying it to specific inspection tasks. This pre-training phase allows the model to learn universal geometric patterns that can be transferred to various component detection scenarios, reducing the need for task-specific training data while maintaining high detection accuracy

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent changes parameters by using geometric transformations (rotation, translation, scaling) to augment the training data. Instead of requiring大量 labeled images of components at various orientations, the system generates synthetic training samples by transforming a limited set of labeled examples, thereby reducing the actual data collection burden while improving model robustness to orientation variations

Inventive Principle:
Principle #35Parameter changes

2Ease of manufacture

If rule-based systems are used for verification, then ease of understanding and implementation is improved, but maintenance difficulty increases as product complexity grows

Engineering Contradiction:
Improveease of implementationVSAvoidrule complexity
Core Design Contradiction:
Ease of manufactureVSDevice complexity

Solution Approach 1:

The patent replaces the mechanical rule-based system with an AI-based perceptual system. Instead of manually coding rules to detect clips, ribs, and bosses, the system uses a Deep Learning model that automatically learns detection patterns from training data. This substitution maintains ease of implementation during deployment while eliminating the maintenance burden of updating complex rules as products evolve

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent applies self-service by enabling the AI model to automatically adapt to new product types through geometric transformation-based data augmentation. When a new component type needs to be detected, the system can generate training samples by transforming existing examples rather than requiring manual rule creation, allowing the system to serve itself in adapting to product complexity changes

Inventive Principle:
Principle #25Self-service

3Difficulty of detecting and measuring

If 3D segmentation or detection network architectures are used to detect components, then component classification and localization capability is improved, but computational complexity and processing time increase

Engineering Contradiction:
Improvecomponent detection capabilityVSAvoidnetwork architecture complexity
Core Design Contradiction:
Difficulty of detecting and measuringVSDevice complexity

Solution Approach 1:

The patent applies segmentation by breaking down the complex task of component detection into a two-stage process: first detecting candidate regions in the 3D point cloud, then classifying and localizing components within those regions. This segmentation of the detection pipeline reduces computational complexity compared to attempting to detect and classify all components in a single pass, while maintaining high detection capability

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4095795B1Automated optical guide
Publication Date: 2024.08.14 FUJITSU LTD
  • EP4095795B1 patent drawingFigure 1
  • EP4095795B1 patent drawingFigure 2
  • EP4095795B1 patent drawingFigure 3

AI summary

A computer-implemented automatic inspection method comprises computing target feature vectors for points in a 3D point-cloud representation of an inspection target, computing component feature vectors for points in a 3D point-cloud representation of a component, and performing matching between the target feature vectors and the component feature vectors to obtain point pairs each comprising a said point in the 3D point-cloud representation of the inspection target and a said point in the 3D point-cloud representation of the component to determine whether the component is present in the inspection target.