Real-Time Pose Detection Using Nodal Point Feature Descriptors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing pose detection methods require significant computational resources and high frame rates, failing to accurately predict poses if frame rates are low or preceding frames are unavailable.
Innovation Solution
A method and system that identify nodal points in an input image, compute feature descriptors based on these points, and compare them with pre-trained descriptors using deep learning techniques to detect poses in real-time, allowing for accurate pose detection with two-dimensional image frames and reduced computational efforts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If training-based models are used for pose detection, then pose detection accuracy is improved, but computational resources and training time requirements increase significantly
Solution Approach 1:
The patent pre-computes and stores feature descriptors for multiple predefined poses during an offline training phase. During real-time operation, the system only needs to compare the computed feature descriptor of the current pose against these pre-stored descriptors, eliminating the need for complex real-time training computations and significantly reducing computational resource requirements while maintaining accuracy.
Solution Approach 2:
The patent creates a library of pre-computed feature descriptors that copy the essential pose information from training data. Instead of performing complex pose estimation computations during inference, the system uses these copied feature descriptors to quickly identify poses through comparison, reducing real-time computational burden while preserving detection accuracy.
2Measurement precision
If high frame rate inputs and preceding frames are required, then pose prediction accuracy is improved, but system responsiveness and real-time performance deteriorate
Solution Approach 1:
The patent extracts and uses only the current input image frame to compute the feature descriptor, completely removing the dependency on preceding frames or high frame rate inputs. The pose detection is performed independently on each frame using the extracted nodal points and computed feature descriptor, enabling real-time processing even with low frame rates or missing frames while maintaining prediction accuracy.
3Adaptability or versatility
If existing pose detection approaches are used, then pose detection capability is provided, but adaptability to low frame rates and missing frames is reduced
Solution Approach 1:
The patent changes the fundamental parameter of how pose information is represented and processed. Instead of using temporal sequences of frames or complex temporal models, the system transforms the pose detection problem into a spatial feature descriptor comparison problem. This parameter change allows the system to operate reliably with any frame rate by focusing on the instantaneous pose representation through nodal points and feature descriptors.
Data Source
AI summary
Disclosed herein is method and system for detecting pose of a subject in real-time. In an embodiment, nodal points corresponding to the subject may be identified and used for identifying skeleton pose of the subject. Thereafter, a feature descriptor for the skeleton pose may be computed based on the nodal points. Further, the feature descriptor of the skeleton pose may be compared with predetermined feature descriptors for detecting the pose of the subject as predefined pose corresponding to one of the predetermined feature descriptors used for the comparison. The method of present disclosure makes accurate pose detection from a two-dimensional image of the subject, using a pose detection model, which is pre-trained with predetermined feature descriptors and deep learning techniques.