3D Hand Keypoint Gesture Recognition Across Viewing Angles
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current gesture recognition methods based on RGB images and hand key points suffer from low recognition accuracy due to the impact of varying viewing angles, leading to misrecognition issues.
Innovation Solution
A gesture recognition method that determines relative and absolute position feature vectors of hand region feature points, using multi-layer perceptrons to process these vectors and account for viewing angle variations, thereby improving recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If gesture recognition is performed based on original RGB image or hand key point modeling, then the recognition process can be implemented, but the recognition accuracy is low due to the impact of varying viewing angles
Solution Approach 1:
The patent transforms the hand key point coordinates from image coordinate system to three-dimensional spatial coordinate system, changing the parameter representation from two-dimensional (x, y) to three-dimensional (x, y, z). This parameter transformation enables the system to account for viewing angle variations by incorporating depth information, thereby improving gesture recognition accuracy across different viewing angles
Solution Approach 2:
The patent introduces a third dimension (depth/z-axis) to the traditional two-dimensional hand key point coordinates. By converting image plane coordinates (x, y) into spatial coordinates (x, y, z) using depth information from structured light or time-of-flight data, the system creates a three-dimensional hand model that is invariant to viewing angle changes, thus resolving the accuracy issue caused by varying perspectives
Data Source
AI summary
A gesture recognition method is provided, including: obtaining a sub-image of a hand region in a target image, and determining multiple pieces of feature point position information corresponding to multiple feature points in the sub-image of the hand region; determining a first position feature vector based on the multiple pieces of feature point position information, where the first position feature vector represents a relative position relationship of any one of the multiple feature points relative to remaining feature points in the multiple feature points; determining a second position feature vector based on the multiple pieces of feature point position information, where the second position feature vector represents an absolute position relationship of the multiple feature points in the sub-image of the hand region; and outputting a recognition result of the sub-image of the hand region based on the first position feature vector and the second position feature vector.