High-Speed Tracking with Kernelized Correlation Filters.

Area of Science:

• Computer Vision
• Machine Learning
• Signal Processing

Background:

• Modern image trackers rely on discriminative classifiers trained with redundant sample patches.
• Natural image variations necessitate robust training methods to handle scale and translation changes.

Purpose of the Study:

• To develop a more efficient analytic model for training image tracking classifiers.
• To reduce computational and storage complexity in visual tracking algorithms.
• To introduce novel, high-performance tracking methods based on this model.

Main Methods:

• Proposed an analytic model for datasets of translated patches, identifying the data matrix as circulant.
• Utilized the discrete Fourier transform for diagonalization, drastically reducing storage and computation.
• Derived a new kernelized correlation filter (KCF) and a dual correlation filter (DCF) for multi-channel tracking.

Main Results:

• The proposed model reduces storage and computation by orders of magnitude.
• The kernelized correlation filter (KCF) achieves linear complexity for kernel regression.
• Dual correlation filter (DCF) offers a fast, multi-channel extension of linear correlation filters.
• KCF and DCF outperform state-of-the-art trackers (Struck, TLD) on a benchmark dataset.
• Achieved high-speed tracking (hundreds of frames-per-second) with concise implementations.

Conclusions:

• The developed analytic model and derived algorithms (KCF, DCF) offer significant improvements in visual tracking efficiency and performance.
• These methods provide a foundation for faster, more accurate, and computationally lighter object tracking systems.
• The open-source release of the tracking framework encourages further research and development in the field.