Face Tracking via Feature Embeddings for Occlusion Handling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current face tracking methods in video streams are susceptible to changes in illumination, spatial relationships, and occlusions, leading to inefficiencies and confusion between similar-looking individuals, especially in crowded scenes and when subjects leave or re-enter the field of view.
Innovation Solution
A robust face tracking method using features trained for face recognition, employing similarity matching with embeddings or binarized signatures, which allows for tracking individuals even when their appearance changes, and can distinguish between similar-looking persons by integrating location and time attributes into the signature comparison process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional frame-by-frame pixel matching is used for face tracking, then the system is simple to implement, but it is highly susceptible to illumination changes, spatial relationship changes, and occlusions
Solution Approach 1:
The patent transforms the tracking approach by changing the parameter space from raw pixel intensity matching to feature-based representation. Face recognition models extract key features (landmarks, texture patterns, geometric relationships) that are invariant to illumination and spatial changes, thereby improving tracking robustness while managing complexity through parameter transformation
Solution Approach 2:
The patent replaces the mechanical pixel-by-pixel comparison mechanism with a learned feature representation system. Deep learning models automatically learn discriminative features that capture essential face characteristics, substituting the simple but fragile pixel matching mechanism with a more robust feature-based approach
2Reliability
If feature-based face recognition is used for tracking, then tracking robustness to appearance changes is improved, but computational complexity increases
Solution Approach 1:
The patent segments the face tracking problem into distinct components: face detection, feature extraction, signature generation, and matching. By dividing the computational task into modular segments, the system can apply optimized algorithms to each stage, improving overall tracking accuracy while managing computational power requirements through targeted processing
Solution Approach 2:
The patent performs preliminary feature extraction and signature generation in advance, creating compact representations of face characteristics before the actual tracking matching occurs. This preliminary action reduces the computational burden during real-time tracking by working with pre-processed feature signatures rather than raw pixel data
3Measurement precision
If similarity matching with embeddings is used, then the ability to distinguish similar-looking persons is improved, but the complexity of signature comparison increases
Solution Approach 1:
The patent introduces face signatures as an intermediary representation between raw pixel data and identification decisions. These signatures serve as compact, discriminative intermediaries that capture essential face characteristics, enabling precise distinction between similar-looking persons while simplifying the comparison process through structured feature representation
Data Source
AI summary
Systems, and method and computer readable media that store instructions for features matching based tracker.