Consistently Sampled Correlation Filters with Space Anisotropic Regularization for Visual Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Existing correlation filter trackers often use fixed patches and cyclic shifts, leading to unreliable training samples and inconsistencies.
• These methods overlook the discrepancies between training and detection samples, impacting tracking accuracy.

Purpose of the Study:

• To propose and evaluate a consistently sampled correlation filter with space anisotropic regularization (CSSAR).
• To address issues of training sample redundancy, inconsistencies, and drift caused by occlusion in object tracking.

Main Methods:

• Developed a spatiotemporally consistent sampling strategy to reduce training sample redundancy and eliminate inconsistencies.
• Introduced space anisotropic regularization to constrain the correlation filter and mitigate occlusion-induced drift.
• Implemented an optimization strategy using the Gauss-Seidel method for efficient online learning.

Main Results:

• The proposed CSSAR tracker demonstrated superior performance compared to state-of-the-art trackers.
• Evaluations were conducted on standard object tracking benchmarks (OTBs).
• Both qualitative and quantitative results confirmed the tracker's effectiveness.

Conclusions:

• CSSAR offers a robust and efficient solution for object tracking, outperforming existing methods.
• The approach effectively handles training sample inconsistencies and occlusion challenges.
• This work advances the field of correlation filter-based visual tracking.