Pose Recognition Model Using Paired Body-Part Sum and Difference

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing techniques face difficulties in accurately recognizing specific actions, such as 'holding an object with the right hand' and 'holding an object with the left hand', as the same actions using machine learning models, due to fluctuations in training data and increased processing loads.

Innovation Solution

An information processing method that involves identifying positional information about paired body parts, generating a sum and difference component from this information, and training a model to output posture information, using a neural network or mathematical expression.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If traditional machine learning models are trained to recognize specific actions with left and right body parts, then the model can distinguish between different actions, but it becomes difficult to recognize symmetric actions (e.g., holding an object with left hand vs. right hand) as the same action

Engineering Contradiction:
Improveaction recognition accuracyVSAvoidability to recognize symmetric actions as same action
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent transforms symmetric body part coordinates into asymmetric features by calculating differences between paired body parts (e.g., left hand position minus right hand position). This asymmetry transformation allows the model to distinguish between symmetric actions while still recognizing them as the same action category, resolving the contradiction between measurement precision and adaptability.

Inventive Principle:
Principle #4Asymmetry

Solution Approach 2:

The patent changes the parameter representation from absolute coordinates to relative differences between paired body parts. By using difference values instead of absolute positions, the model can generalize across symmetric variations while maintaining accurate action recognition, thus improving both measurement precision and adaptability to symmetric actions.

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If more training data is used to improve model accuracy for specific action recognition, then recognition accuracy improves, but processing load and training costs increase

Engineering Contradiction:
Improveaction recognition accuracyVSAvoidtraining efficiency
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent extracts only the essential features needed for action recognition by calculating differences between paired body parts, rather than using all raw coordinate data. This feature extraction reduces the dimensionality of training data while preserving the most informative aspects, thereby improving training efficiency without sacrificing recognition accuracy.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent segments the body into paired parts (left-right pairs) and processes each pair independently by calculating their differences. This segmentation approach simplifies the overall data structure and reduces the total amount of training data needed, while maintaining the ability to accurately recognize actions through the segmented feature differences.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4379680B1Information processing program, information processing method, and information processing device
Publication Date: 2026.03.18 FUJITSU LTD
  • EP4379680B1 patent drawingFigure 1
  • EP4379680B1 patent drawingFigure 2
  • EP4379680B1 patent drawingFigure 3

AI summary

An information processing program comprising instructions which when executed by a computer, cause the computer to execute processing including: acquiring a video image in which a first person appears; identifying positional information about each of two body parts that form a pair among body parts of the first person in the video image, by analyzing the acquired video image; generating a first component that indicates a sum of the identified positional information about each body part; generating a third component that indicates an absolute value of a second component that indicates a difference in the identified positional information about each body part; and training a model that outputs posture information about a second person from a video image in which the second person appears, based on the generated first component and the generated third component.