Efficient and Practical Correlation Filter Tracking.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Visual tracking is crucial but computationally intensive, limiting applications in resource-constrained environments.
• Existing correlation filter trackers struggle with target deformation and occlusion due to simple update schemes.

Purpose of the Study:

• To develop an efficient and adaptive training sample update scheme for correlation filter-based visual trackers.
• To enhance tracker robustness against target deformation and occlusion.
• To extend the tracker for reliable long-term tracking capabilities.

Main Methods:

• Proposed an efficient and adaptive training sample update scheme using difference hashing for sample selection.
• Implemented a tracking state discrimination mechanism for failure detection and recovery.
• Expanded the tracker for long-term tracking scenarios.

Main Results:

• The proposed tracker demonstrates favorable performance against state-of-the-art methods on benchmark datasets (OTB-2015, Temple Color 128, UAV123).
• Achieved high tracking speeds exceeding 100 fps on a standard desktop CPU.
• Showcased robustness against target deformation and occlusion.

Conclusions:

• The developed adaptive update scheme significantly improves visual tracking efficiency and robustness.
• The tracker is suitable for real-time applications, even in challenging conditions.
• The long-term tracking extension provides reliable target recovery after failure.