Real-Time Gesture Recognition Using Angular Radial Bin Descriptors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current shape matching algorithms for gesture recognition, such as those used in computer vision, face challenges with rotation, translation, scale invariance, and deformation, while also being computationally inefficient, particularly for highly articulated shapes like hand gestures.
Innovation Solution
The method employs Angular-Radial Bin (ARB) distribution-based shape descriptors, which calculate weights for contour pixels within ARBs, creating a histogram that is invariant to rotation, translation, and scale, and uses multiple angle-titled instances of ARB distributions to enhance matching accuracy and efficiency by comparing these descriptors across a database.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If shape context method is used for hand gesture recognition, then accuracy is improved, but time complexity increases to O(n^3)
Solution Approach 1:
The patent segments the shape matching problem by dividing the gesture contour into multiple angular radial bins (ARBs). Each bin independently captures local shape features in specific angular and radial directions, transforming a single complex O(n^3) matching problem into multiple simpler O(1) histogram comparisons. This segmentation enables linear time complexity while maintaining accuracy through distributed feature representation.
Solution Approach 2:
The patent introduces angular and radial dimensions to represent shape features, creating a two-dimensional histogram space (angular dimension × radial dimension). This dimensional transformation converts point-based shape context features into binned histogram features, enabling efficient comparison through histogram intersection or distance metrics rather than complex point-to-point matching, thus reducing time complexity from cubic to linear.
2Reliability
If rotation invariance is implemented in shape matching, then robustness is improved, but computational complexity increases
Solution Approach 1:
The patent pre-computes shape descriptors for multiple rotated instances of the gesture shape before comparison. By generating descriptors for shapes rotated at various angles in advance, the system eliminates the need for real-time rotation handling during matching. This preliminary action ensures rotation invariance while keeping the actual matching process computationally simple, as it only requires comparing pre-computed histograms.
3Measurement precision
If multiple angle tilted instances of ARB distribution are used, then matching accuracy is improved, but computational load increases
Solution Approach 1:
The patent merges multiple angle-titled ARB distributions by accumulating their histogram values into a single composite histogram. Instead of performing separate matching operations for each angle, the method combines the descriptors from different angular perspectives by summing or averaging their histogram bins. This merging strategy maintains the accuracy benefits of multi-angle analysis while reducing computational load by performing a single consolidated comparison.
Data Source
Figure 1
Figure 2~3
Figure 4
AI summary
The present disclosure relates to a method and a device (20) for recognizing hand gestures in real-time. A shape is given as an input in a form of a binary image (14). The shape contour is partitioned into radial and angular spaces (110) by means of an Angular Radial Bin distribution (100) comprising multiple concentric circles (112, 114) and angular space partitions (116) in a way that multiple angular radial sections are created denoted Angular Radial Bins (110). The ARB distribution (100) is angle tilted (5) through its centre of mass multiple times and the same procedure is repeated in order to capture a shape descriptor from different angle perspectives. A shape descriptor is calculated for each of an angle tilted instance (δ, 2δ, 3δ,...) of the ARB distribution belonging to a sequence (n = 0, 1, 2,...,Ν-1) of angle tilted instances (n) of the ARB distribution (100).