Image Processing Apparatus for Accurate Posture Estimation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for estimating a person's posture from video images, such as those captured by monitoring cameras, face challenges in accuracy when the camera is inclined or when a person is in a fallen-down posture, and struggle to distinguish between similar postures based on joint point positions.

Innovation Solution

An image processing apparatus that includes a human body detection unit, image generation unit, joint point detection unit, and posture estimation unit, which generates and processes joint point likelihood maps by rotating and resizing images to align spatial information, and adjusts clip regions for accurate posture estimation, considering both joint points and peripheral objects.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If joint point detection is performed on original images without spatial alignment, then detection speed is maintained, but posture estimation accuracy deteriorates when cameras are inclined or persons are in fallen postures

Engineering Contradiction:
Improveposture estimation accuracyVSAvoidimage processing complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system performs preliminary spatial alignment by generating joint point likelihood maps that are pre-adjusted to match the coordinate system of the posture estimation model. This preliminary transformation ensures that when joint points are detected, they are already in the correct spatial reference frame, eliminating the need for complex real-time coordinate transformations during posture estimation.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent introduces joint point likelihood maps as an intermediary representation between raw image data and posture estimation. These maps serve as a bridge that transforms joint point detection results into a format compatible with posture estimation models, handling spatial transformations in an intermediate stage rather than directly between detection and estimation modules.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Measurement precision

If multiple processing steps including rotation and resizing are performed to align spatial information, then posture estimation accuracy improves, but processing time increases

Engineering Contradiction:
Improveposture estimation accuracyVSAvoidprocessing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

Spatial alignment transformations (rotation, resizing) are performed preliminarily when generating joint point likelihood maps, so that the coordinate systems are aligned before posture estimation begins. This preliminary preparation eliminates the need for repeated transformations during the actual posture estimation process, reducing overall processing time.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system performs spatial alignment only for the specific regions and scales necessary for posture estimation, rather than processing entire images at multiple resolutions. By focusing transformations only where needed for accurate joint point detection and posture estimation, processing overhead is minimized while maintaining accuracy.

Inventive Principle:
Principle #16Partial or excessive action

3Reliability

If clip regions are adjusted to include peripheral objects, then context information for posture classification improves, but region of interest precision may deteriorate

Engineering Contradiction:
Improveposture classification reliabilityVSAvoidjoint point detection precision
Core Design Contradiction:
ReliabilityVSMeasurement precision

Solution Approach 1:

The system segments the image processing into different functional regions: a primary region of interest for accurate joint point detection, and extended peripheral regions for context gathering. By dividing the processing area into these segments, the system can maintain high precision in the core detection region while incorporating contextual information from surrounding areas for improved posture classification.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Different quality levels are applied to different regions: high-resolution processing and strict precision requirements are applied to the central region containing the person, while lower-resolution processing with more flexible parameters is applied to peripheral regions used for context. This local quality differentiation maintains detection precision where needed while gathering sufficient contextual information.

Inventive Principle:
Principle #3Local quality

Data Source

PatentEP4459571A1Image processing apparatus, method of controlling image processing apparatus, program, and storage medium
Publication Date: 2024.11.06 CANON KK
  • EP4459571A1 patent drawingFigure 1
  • EP4459571A1 patent drawingFigure 2
  • EP4459571A1 patent drawingFigure 3

AI summary

An image processing apparatus (100) includes human body detection means (202) for detecting a region of a human body from an image, image generation means (203) for generating, from the image, a first clipped image (501) including the region of the human body detected by the human body detection means (202) and a second clipped image (502) different from the first clipped image (501) and including the region of the human body, joint point detection means (204) for detecting joint points of the human body from the first clipped image (501) to generate joint point information (602) on the human body, conversion means (205) for converting the joint point information (602) so as to have spatial information coincident with spatial information on the second clipped image (502), and posture estimation means (206) for estimating a posture of the human body based on the joint point information (602) converted by the conversion means (205) and the second clipped image (502).