Robust deformable and occluded object tracking with dynamic graph.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Deformation and occlusion remain significant challenges in visual object tracking.
• Existing methods often struggle to robustly handle these issues in real-world scenarios.

Purpose of the Study:

• To propose a unified framework for robust visual tracking that addresses both deformation and occlusion.
• To introduce a novel dynamic graph-based tracker (DGT) for enhanced tracking performance.

Main Methods:

• Representing targets as dynamic graphs with nodes for local parts and edges for geometric structure.
• Formulating tracking as a graph matching problem between target and candidate graphs.
• Utilizing Markov Random Fields for background separation and spectral clustering for graph matching.
• Employing weighted voting and foreground/background segmentation for state determination and refinement.
• Implementing an online updating mechanism for adaptive model adjustments.

Main Results:

• The dynamic graph representation captures richer target information, crucial for handling deformation and occlusion.
• The DGT framework demonstrated superior performance compared to several state-of-the-art trackers.
• The tracker showed robustness across various challenging visual tracking scenarios.

Conclusions:

• The proposed dynamic graph-based tracker offers a robust and unified solution for visual tracking under deformation and occlusion.
• The DGT framework effectively adapts to target structure variations, enhancing tracking reliability.