State-aware Anti-drift Object Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Correlation filter (CF) trackers offer high speed and performance in visual tracking.
• Existing trackers struggle with model drift due to target interference and distractions.
• Robustness against external and internal interference is crucial for reliable tracking.

Purpose of the Study:

• To develop a State-aware Anti-drift Tracker (SAT) that improves tracking robustness.
• To jointly model discrimination and reliability information for enhanced filter learning.
• To mitigate model drift in visual tracking under challenging conditions.

Main Methods:

• Incorporated global context patches for improved target-background discrimination.
• Learned a color-based reliable mask to focus on robust tracking regions.
• Utilized the Alternative Direction Method of Multipliers (ADMM) for efficient optimization in the Fourier domain.
• Implemented a Kurtosis-based updating scheme for high-confidence template updates.

Main Results:

• The proposed SAT tracker demonstrates superior performance on OTB-100 and UAV-20L datasets.
• Quantitative and qualitative evaluations confirm the effectiveness and robustness of SAT.
• The method successfully distinguishes targets from backgrounds and resists model drift.

Conclusions:

• The State-aware Anti-drift Tracker (SAT) offers a significant advancement in visual tracking.
• Jointly modeling discrimination and reliability enhances tracker resilience.
• SAT provides a robust and efficient solution for real-world visual tracking applications.