Efficient Correlation Tracking via Center-Biased Spatial Regularization.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Correlation filters (CFs) offer high accuracy and efficiency in visual tracking.
• The periodic assumption in CFs using Fast Fourier Transform (FFT) causes boundary effects.
• Spatially-Regularized Discriminative CF (SRDCF) improves accuracy but increases computational complexity due to lack of closed-form FFT solution.

Purpose of the Study:

• To develop an efficient and effective CF-based tracker with improved speed and accuracy.
• To address the computational complexity and boundary effects of existing CF methods.
• To enhance robustness in visual tracking, particularly during occlusions.

Main Methods:

• A novel center-biased constraint weights (CBCW) function is constructed using Fourier transform symmetry.
• CBCWs are real in both time and frequency domains, enabling direct frequency-domain optimization.
• An efficient filter update strategy based on average peak-to-correlation energy is proposed for occlusion handling.

Main Results:

• The proposed tracker significantly outperforms the baseline SRDCF in both accuracy and efficiency.
• Experiments on OTB-2013, OTB-2015, and VOT2016 benchmarks validate the tracker's performance.
• The method demonstrates favorable robustness and success rates compared to 16 other state-of-the-art trackers.

Conclusions:

• The proposed CF tracker with CBCWs offers a superior balance of speed and accuracy.
• The CBCW approach effectively reduces computational complexity and mitigates boundary effects.
• The developed tracker represents a significant advancement in efficient and robust visual tracking.