Posture Estimation Using Joint Consistency and Dynamic Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing posture estimation methods for multi-joint objects are inconsistent, particularly when dealing with objects of changing shape, as they rely on fixed local feature arrangements and lack consideration for local portions, making them unsuitable for non-rigid bodies.
Innovation Solution
A posture estimating apparatus that inputs range images, derives joint position candidates, and estimates posture based on positional relations between joints in a multi-joint object model, using a unit configured to input range images, calculate joint positions, and evaluate consistency with a stored human body model to determine a consistent posture.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If fixed local feature arrangement methods are used for posture estimation, then the method works well for rigid bodies with predetermined shapes, but it cannot be applied to objects with highly changed shapes
Solution Approach 1:
The patent applies dynamics by transitioning from fixed, static posture models to dynamic models that adapt to changing object shapes. The system uses multiple posture candidates with different confidence levels and dynamically selects the most appropriate posture based on current image data, allowing the estimation to flex with varying object configurations rather than relying on predetermined rigid structures
Solution Approach 2:
The patent changes parameters by introducing confidence levels and probability distributions for different posture candidates. Instead of relying on a single fixed posture model, the system maintains multiple posture representations with varying confidence values, allowing it to adapt to different object shapes by selecting postures with higher confidence levels for the current observation
2Productivity
If the whole posture is uniquely determined based on a previously set model, then the possible posture of the object can be obtained, but consistency of the posture including local portions is insufficient
Solution Approach 1:
The patent applies segmentation by dividing the posture estimation into multiple independent candidate postures, each representing a different possible configuration. Instead of determining a single whole posture, the system segments the solution space into multiple discrete posture candidates with associated confidence levels, then selects the most appropriate segment (posture) based on current observations
Solution Approach 2:
The patent implements feedback by using confidence levels associated with each posture candidate to guide the selection process. The system evaluates local portion consistency as feedback to adjust the selection among posture candidates, ensuring that the final determined posture maintains consistency across all local portions while still achieving efficient determination
3Reliability
If multiple joint position candidates are derived for each joint position, then more posture possibilities are considered, but the complexity of evaluating all combinations increases
Solution Approach 1:
The patent applies partial action by not evaluating all possible combinations of joint position candidates exhaustively. Instead, it derives multiple candidates for each joint but selectively evaluates only those combinations that are most likely to represent the actual posture, using confidence levels and local consistency checks to prune the search space and avoid unnecessary computational complexity
Data Source
AI summary
The present invention aims to estimate a more consistent posture in regard to a multi-joint object. A target range image is first input, a human body region is extracted from the input range image, a target joint position candidate is calculated from the input range image, and a joint position is finally determined based on the calculated joint position candidate and a likelihood of each joint to estimate the posture. At this time, joint position permissible range information concerning inter-joint distance and angle of a human body model previously set by learning is obtained from a human body model storing unit, consistency is evaluated for a relation between the joint position candidates of a certain joint and other joint based on the obtained information, and thus the posture corresponding to the best combination of the joint positions is determined.