Multi-view 3D positioning via geometric clustering

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing techniques for 3D positioning and pose determination in computer vision face challenges such as reliance on specific object types, sensitivity to occlusion and crowding, and inefficiency in processing, particularly when dealing with multiple objects and noisy image data.

Innovation Solution

A method that processes object detection data from multiple views to compute 3D positions and poses by generating candidate association data, clustering estimated positions, and determining final associations, which is robust to inaccuracies and applicable to any type of object without relying on predefined feature databases or object-specific heuristics.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If appearance features are used for 3D pose determination, then the technique works well for human pose, but it fails for objects with similar appearance and requires large training datasets

Engineering Contradiction:
Improvepose determination accuracyVSAvoidapplicability to different object types
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent applies universality by using geometric features that are common to all rigid objects rather than appearance-specific features. The method determines 3D pose by identifying corresponding geometric features (edges, corners, surfaces) across multiple views and applying rigid transformation mathematics, making it applicable to any rigid object regardless of appearance, thus resolving the contradiction between accuracy and versatility.

Inventive Principle:
Principle #6Universality (Multi-functionality)

2Measurement precision

If feature databases are prepared for multiple object types and poses, then object recognition accuracy improves, but the complexity of the system increases

Engineering Contradiction:
Improveobject recognition accuracyVSAvoidfeature database management
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent applies self-service by enabling the system to automatically determine 3D pose from geometric correspondences without requiring pre-prepared feature databases. The method computes pose directly from detected geometric features in multiple views using rigid transformation equations, eliminating the need for manual database creation and management while maintaining high recognition accuracy.

Inventive Principle:
Principle #25Self-service

3Productivity

If greedy algorithms are used for the second stage processing, then processing efficiency improves, but robustness to detection inaccuracies decreases

Engineering Contradiction:
Improveprocessing speedVSAvoidrobustness to detection errors
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent applies feedback by using all detected geometric features from multiple views to compute 3D pose through rigid transformation equations, rather than making greedy local optimizations. The method iteratively refines pose estimates by checking consistency across all feature correspondences, providing feedback that corrects detection errors and improves robustness while maintaining processing efficiency through direct mathematical computation.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP3836085B1Multi-view three-dimensional positioning
Publication Date: 2024.06.12 SONY GROUP CORP
  • EP3836085B1 patent drawingFigure 1A~1D
  • EP3836085B1 patent drawingFigure 2A~2C
  • EP3836085B1 patent drawingFigure 2D~2G

AI summary

A device determines the positioning of objects in a scene by implementing a robust and deterministic method. The device obtains (201) object detection data (ODD) which identifies the objects and locations of reference points of the objects in 2D images of the scene. The device processes (202) the ODD to generate candidate association data (CAD) which associates pairs of objects between the images, computes (203) estimated 3D positions in the scene for associated pairs of objects in the CAD, and performs clustering (204) of the estimated positions. The device further generates (205), based on estimated 3D positions in one or more clusters, final association data (FAD) which associates one or more objects between the images, and computes (206) one or more final 3D positions in the scene for one or more reference points of the one or more objects in the FAD. The final 3D position(s) thereby represent the 3D position and/or the 3D pose of the respective object in the scene.