Human Pose Prediction Using Graph Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing human pose prediction methods based on image features have limited prediction precision as they primarily focus on mining image features without effectively utilizing spatial information between key-points.
Innovation Solution
A method that invokes a basic neural network to obtain an initial prediction graph, which is then refined using a pose graph neural network to mine spatial information, resulting in a final prediction graph that combines image feature extraction with spatial information between human key-points, achieving higher accuracy in human pose estimation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a prediction method based on image features is used, then the method focuses on mining image features, but the prediction precision is limited
Solution Approach 1:
The patent combines image feature extraction with spatial relationship modeling by integrating a pose graph neural network with a convolutional neural network. The pose graph neural network processes spatial relationships between key points while the convolutional neural network extracts image features, merging both approaches to achieve higher prediction precision without excessive complexity increase
Solution Approach 2:
The patent transitions from two-dimensional image feature extraction to three-dimensional spatial relationship analysis by introducing pose graph neural networks that model temporal and spatial dependencies between key points across multiple frames, adding a temporal dimension to the traditional spatial analysis
2Measurement precision
If iterative prediction is performed using pose graph neural network, then spatial information between key-points is mined, but computation time increases
Solution Approach 1:
The patent performs preliminary actions by pre-processing images to extract initial key point positions and pre-building pose graphs before the main iterative prediction process. This preparation reduces the computational burden during iterative refinement by having the structural framework already in place
Solution Approach 2:
The patent implements feedback mechanisms where the pose graph neural network iteratively refines predictions by comparing predicted key point positions with ground truth and adjusting accordingly. The feedback loop allows the system to converge to accurate solutions more efficiently by learning from prediction errors rather than requiring exhaustive computation
Data Source
Figure 1~3
Figure 4
Figure 5
AI summary
Disclosed by embodiments of the present application are a body posture prediction method, an apparatus, a device, and a storage medium, which belong to the field of image processing. The method comprises: calling a basic neural network to predict an inputted target image, and obtaining an initial predicted image of key points of the body; inputting the initial predicted image of key points of the body and a body structure image into a body image neural network; calling the body image neural network using the initial predicted image of key points of the body to initialize the body structure image, and obtaining an initialized body structure image; calling the body image neural network to iteratively predict the initialized body structure image, and obtaining a final predicted image. The embodiments of the present application are capable of fully utilizing a body image neural network to mine spatial information between joints of the body, thus comprehensively mining spatial information between the joints of the body on the basis of extracting image features, and obtaining body posture information that is more accurate.