Video Tracking via Key Point Posture Geometry
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing person tracking technologies fail to accurately track multiple individuals across video frames, especially when conditions such as congestion, view angle, distance, and frame rate differ from learned data, and require reference images for each posture stored in a database.
Innovation Solution
A tracking apparatus that detects targets in video frames, extracts key points, generates posture information, and tracks targets based on position and orientation across frames, allowing continuous tracking without the need for reference images or specific learning conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning-based posture tracking is used, then tracking accuracy under learned conditions is improved, but tracking reliability under varying conditions (congestion, angle, distance, frame rate) deteriorates
Solution Approach 1:
The patent changes the tracking parameters from deep learning model predictions to geometric relationships between key points. By representing posture as relative positions and orientations of key points rather than relying on learned patterns, the system adapts to varying conditions (congestion, angle, distance, frame rate) without requiring retraining or reference images.
2Measurement precision
If reference images for each posture are stored in a database, then person identification accuracy is improved, but device complexity and storage requirements increase
Solution Approach 1:
The patent extracts only the essential geometric features (key point positions and orientations) needed for tracking, eliminating the need to store complete reference images. By taking out only the necessary positional and orientational data rather than storing full image references, the system reduces storage complexity while maintaining tracking capability.
Solution Approach 2:
Instead of storing actual reference images, the patent creates simplified geometric copies represented by key point coordinates and relative orientations. These geometric representations serve as lightweight substitutes that capture the essential posture information without the complexity of storing full image data.
3Measurement precision
If three-dimensional posture estimation from two-dimensional joint positions is performed, then posture information quality is improved, but ability to track multiple persons deteriorates
Solution Approach 1:
The patent segments the tracking problem by independently tracking key points for each person rather than estimating complete three-dimensional postures. By dividing the multi-person tracking task into independent key point tracking segments, the system maintains posture information quality while scaling to multiple persons without the computational complexity of full 3D estimation for each individual.
Data Source
AI summary
A tracking apparatus that includes a detection unit that detects a tracked target from at least two frames constituting video data; an extraction unit that extracts at least one key point from the tracked target having been detected, a posture information generation unit that generates posture information of the tracked target based on the at least one key point, and a tracking unit that tracks the tracked target based on a position and an orientation of the posture information of the tracked target detected from each of the at least two frames.