Scene-Aware Adaptive Updating for Visual Tracking via Correlation Filters.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Visual object tracking is crucial for applications like surveillance and human-computer interaction.
• Correlation filter-based trackers offer speed and accuracy but struggle with scale variation, deformation, and occlusion.

Purpose of the Study:

• To enhance the robustness of kernel correlation filter (KCF) based visual object tracking.
• To address limitations of existing trackers in challenging video scenes.

Main Methods:

• A scene-aware adaptive updating mechanism for KCF visual tracking is proposed.
• A low-complexity scale estimation method using weighted five scales is introduced.
• Video scenes are classified into four categories for adaptive filter updates.

Main Results:

• The proposed tracker shows significant performance improvements over the standard KCF tracker on the CVPR2013 benchmark.
• Improvements include 33.3% for scale variation, 15% for occlusion, 6% for deformation, 21.9% for fast motion, and 19.8% for out-of-view scenarios.

Conclusions:

• The scene-aware adaptive updating mechanism effectively improves visual object tracking robustness.
• The method demonstrates superior performance in complex scenes compared to traditional KCF trackers.