Segmented Pose Estimation Networks for Resource-Limited Devices

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing human pose estimation algorithms are not suitable for devices with limited computing capacity, leading to low accuracy and inability to support applications effectively.

Innovation Solution

A lightweight neural network structure is developed by adjusting the block and bottleneck units of the feature extraction and backend prediction neural networks, integrating them to reduce computational load while maintaining high accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If existing human pose estimation algorithms are used, then pose estimation capability is provided, but the algorithms cannot be implemented on devices with limited computing capacity

Engineering Contradiction:
Improvecompatibility with terminal devicesVSAvoidcomputational resource requirements
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The neural network is divided into two separate modules: a backbone network for feature extraction and a prediction network for pose estimation. This segmentation allows each module to be optimized independently, reducing the overall computational burden on terminal devices while maintaining pose estimation functionality.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent transforms the traditional single-network architecture into a modular two-network architecture, adding a dimensional structure to the system. This dimensional change enables better resource management and computational efficiency on constrained devices.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Measurement precision

If existing human pose estimation algorithms are implemented on terminal devices, then pose estimation is performed, but the accuracy is insufficient to meet application requirements

Engineering Contradiction:
Improvepose estimation accuracyVSAvoidcomputational efficiency
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

By separating feature extraction and pose prediction into distinct networks, each can be specialized and optimized for its specific task. The backbone network focuses on extracting meaningful features efficiently, while the prediction network specializes in accurate pose estimation, achieving both speed and accuracy.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The backbone network serves multiple purposes: it extracts features for pose estimation and can be reused across different pose estimation scenarios. This multi-functionality improves computational efficiency without compromising accuracy.

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

3Measurement precision

If high-performance pose estimation algorithms are used, then accurate pose estimation is achieved, but the algorithms cannot be supported by terminal devices with limited computing capacity

Engineering Contradiction:
Improvepose estimation accuracyVSAvoidcomputational energy consumption
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The divided architecture allows the backbone network to be computationally efficient for mobile devices while the prediction network maintains high accuracy requirements. This segmentation enables accurate pose estimation with reduced energy consumption on terminal devices.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP3740936B1Method and apparatus for pose processing
Publication Date: 2025.07.30 SAMSUNG ELECTRONICS CO LTD
  • EP3740936B1 patent drawingFigure 1~3A
  • EP3740936B1 patent drawingFigure 3B~4A
  • EP3740936B1 patent drawingFigure 4B~7

AI summary

Provided is a method for pose estimation in a device, the method comprising capturing an image; estimating poses of an object included in the captured image; obtaining skeleton information of the object based on the estimating of the poses of the object; and processing the skeleton information of the object for at least one of detecting blocking of the object, detecting the poses of the object and adjusting content based on detected virtual object distinct from human body poses.