Real-Time Object Tracking via Adaptive Correlation Filters.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Correlation filter-based trackers (CFTs) offer robustness and accuracy but can be limited by single templates or fixed feature fusion.
• Existing trackers struggle with target appearance variations in complex scenarios, leading to performance degradation.

Purpose of the Study:

• To propose a real-time dual-template CFT designed for enhanced performance in challenging tracking scenarios.
• To address limitations of single-template and fixed-feature fusion methods in current trackers.

Main Methods:

• Extraction of color histograms, Histogram of Oriented Gradients (HOG), and Color Naming (CN) features.
• Implementation of a dual-template strategy based on target response confidence.
• Integration of discriminative appearance models, multi-peak target re-detection, and scale adaptation.
• Employment of high-confidence template updating to prevent filter model drift.

Main Results:

• The proposed dual-template CFT demonstrated superior performance compared to state-of-the-art handcrafted features-based trackers (HFTs).
• The tracker achieved performance comparable to deep features-based trackers (DFTs) in comprehensive experiments.
• Evaluated against 27 trackers on four benchmark databases, showing favorable results.

Conclusions:

• The real-time dual-template CFT effectively handles appearance variations and model drift in challenging tracking tasks.
• The proposed method offers a competitive alternative to existing HFTs and DFTs, enhancing tracking robustness and accuracy.