Skeleton Estimation via Cylinder Models and Angle Adjustment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing skeleton estimation methods face challenges in accurately recognizing the skeleton of a user at high speed due to noise from 3D sensors, label partitioning errors, and ambiguous pixel labeling, leading to unstable joint position estimation and prolonged calculation times, especially when dealing with complex human body postures.
Innovation Solution
A skeleton estimation device and method that employs a combination of image thresholding, thinning, feature point extraction, and model verification processes to accurately specify body parts and estimate skeleton positions using cylinder models, with a second estimation unit adjusting angle parameters to improve recognition speed and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If body part recognition method using random forest discriminator is used, then joint coordinates can be directly estimated from label centers of gravity, but the skeleton recognition accuracy deteriorates due to noise from 3D sensors, label partitioning errors, and ambiguous pixel labeling
Solution Approach 1:
The patent introduces cylinder models as intermediary objects between the raw distance image data and the final skeleton estimation. The cylinder models serve as mediators that filter out noise and ambiguities from 3D sensors and pixel labeling, providing stable geometric representations for joint position determination. This intermediary step resolves the contradiction by enabling accurate skeleton recognition without directly relying on noisy label centers of gravity.
Solution Approach 2:
The patent replaces the statistical machine learning approach (random forest discriminator) with a geometric modeling approach using cylinder models. This substitution eliminates the need for complex discriminators while providing more reliable joint position estimation through geometric constraints, thereby improving accuracy without sacrificing calculation speed.
2Measurement precision
If model fitting method with cylinder models is used, then skeleton recognition accuracy is improved through geometric modeling, but calculation time is prolonged due to repeated execution of shape calculation processes
Solution Approach 1:
The patent performs preliminary processing by generating binary images and extracting feature points before the main model fitting process. This preliminary action prepares the data in advance, reducing the computational burden during the repeated shape calculation processes and thereby shortening the overall calculation time while maintaining accuracy.
Solution Approach 2:
The patent segments the skeleton recognition process into distinct stages: binary image generation, feature point extraction, and model fitting. This segmentation allows each stage to be optimized independently, with the feature point extraction stage preparing simplified data that accelerates the subsequent model fitting process, thus reducing total calculation time.
3Stability of the object's composition
If feature points are extracted from binary images and cylinder models are fitted, then joint position estimation is stabilized, but device complexity increases due to multiple processing units and verification processes
Solution Approach 1:
The patent divides the processing system into specialized units (binary image generation unit, feature point extraction unit, model fitting unit, verification unit) that each perform specific functions. This segmentation stabilizes joint position estimation by ensuring each processing stage is dedicated to its specific task, while the modular structure actually reduces overall complexity compared to a monolithic system.
Solution Approach 2:
The verification unit automatically verifies whether extracted feature points correspond to actual joints using the cylinder models, without requiring external intervention. This self-service mechanism stabilizes joint position estimation by filtering out false positives, while maintaining system simplicity through automated verification rather than complex manual validation processes.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A skeleton estimation device (100) roughly specifies the areas of a subject's head, both arms, body, and both legs, and, after recognizing the positions of the skeletons of both arms and both legs, recognizes the skeleton of the body, with the recognition result of the positions of the skeletons of both arms and both legs as a constraint condition. For example, the skeleton estimation device (100) varies an angle parameter in a state in which the coordinates of the skeleton for both arms and both legs and the joint to joint lengths are fixed, and thereby calculates a plurality of candidates for the position of the skeleton of the body. The skeleton estimation device (100) estimates a candidate close to approximate coordinates of the body among the candidates for the position of the skeleton of the body, as the position of the skeleton of the body.