Robust visual tracking using local sparse appearance model and K-selection.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Online tracking algorithms adapt to appearance changes but suffer from error accumulation and drift, particularly in occluded scenarios.
• Combining different tracking approaches can improve robustness by balancing stability and flexibility.

Purpose of the Study:

• To develop a robust object tracking algorithm that mitigates drift issues inherent in online learning methods.
• To enhance tracking performance, especially under challenging conditions like occlusion, by integrating sparse representation and a novel selection mechanism.

Main Methods:

• A local sparse appearance model (SPT) was employed, utilizing both static and dynamically updated dictionaries to represent target appearance.
• A novel sparse representation-based voting map and sparse constraint regularized mean shift were developed for robust object tracking.
• A new K-Selection algorithm was introduced for dictionary learning, incorporating locally constrained sparse representation.

Main Results:

• The proposed algorithm demonstrated superior performance compared to existing state-of-the-art methods in comprehensive experimental evaluations.
• The integration of SPT and K-Selection effectively addressed the drifting problem in online tracking, improving accuracy and stability.
• The algorithm showed robust performance, particularly in handling occluded scenarios.

Conclusions:

• The developed tracking algorithm offers a significant improvement in robustness and accuracy over current methods.
• The K-Selection algorithm and sparse representation techniques provide an effective solution for challenging object tracking tasks.
• This work contributes a more stable and flexible approach to online learned tracking.