Joint Point Detection Using Trunk-Based Region Partitioning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current pose estimation methods using deep learning and Convolutional Neural Networks (CNN) struggle to detect joint points when limbs are at specific angles, leading to undetectable joints, particularly for covered limb joints like elbows and knees.
Innovation Solution
A method and device for joint point detection that involves finding a trunk joint point, defining a region of interest, identifying partitions corresponding to sub-limbs, and updating the human body pose skeleton by defining a specific limb joint point based on these partitions, even if limb joints are covered.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If standard pose estimation methods (top-down or bottom-up) are used, then detection process is simple, but joint points cannot be detected when limbs are at specific angles or covered
Solution Approach 1:
The patent divides the detection process into multiple stages: first detecting trunk joint points, then defining regions of interest around each trunk joint, and finally detecting limb joint points within these regions. This segmentation allows the system to handle covered joints by breaking down the complex detection task into manageable parts, improving reliability without overwhelming complexity
Solution Approach 2:
The patent introduces a hierarchical detection dimension by first detecting trunk joints and then using those positions to define regions of interest for limb joint detection. This two-dimensional approach (trunk → limbs) enables the system to infer covered joint positions through spatial relationships and body structure knowledge, rather than attempting to detect all joints simultaneously in a single pass
2Measurement precision
If regions of interest are defined for each trunk joint point and partitions are identified, then covered limb joints can be detected, but processing time and computational complexity increase
Solution Approach 1:
The patent applies local quality by defining regions of interest only around trunk joint points rather than processing the entire image uniformly. Within each region of interest, the system identifies partitions corresponding to specific body parts (arms, legs). This localized approach concentrates computational resources on areas where joint detection is needed, improving accuracy while reducing overall processing time compared to full-image analysis
Solution Approach 2:
The patent performs preliminary detection of trunk joint points before proceeding to limb joint detection. By first establishing the positions of trunk joints and defining regions of interest around them, the system prepares the detection framework in advance. This preliminary action enables faster subsequent detection of limb joints within the predefined regions, reducing total processing time while maintaining high accuracy
Data Source
AI summary
The disclosure provides a method and a device for joint point detection. The method includes: obtaining a human body image and finding joint points to generate a human body pose skeleton; finding a trunk joint point corresponding to a trunk joint; defining a region of interest corresponding to the trunk joint point in the human body pose skeleton, wherein the trunk joint point is connected to a limb of a human body, the limb includes sub-limbs connected to each other via a limb joint; finding partitions corresponding to the sub-limbs based on the region of interest; defining a specific limb joint point based on the partitions; and defining the specific limb joint as corresponding to the limb joint and accordingly updating the human body pose skeleton.