Reliable Multi-kernel Subtask Graph Correlation Tracker.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Existing correlation filter trackers often lack sufficient focus on tracking reliability and precise localization.
• This limitation hinders their effectiveness in complex visual tracking scenarios.

Purpose of the Study:

• To develop a novel particle filter tracker that significantly improves tracking reliability and locating accuracy.
• To address the shortcomings of current correlation filter-based tracking methods.

Main Methods:

• Proposed a graph regularized multi-kernel, multi-subtask learning framework for cross-correlation particle filtering.
• Employed reliable feature selection and assigned multiple non-linear kernels to multi-channel features.
• Utilized multi-view cooperation and interdependencies of target subregions for joint filter learning, incorporating Laplacian graph regularization for temporal-spatial consistency.

Main Results:

• The learned filters comprise a weighted combination of base kernels and reliable integration of base filters, enhancing feature reliability and excluding distractive elements.
• Demonstrated remarkable and competitive performance compared to state-of-the-art methods across five diverse datasets.

Conclusions:

• The proposed tracker effectively enhances tracking reliability and locating accuracy.
• The graph regularized multi-kernel multi-subtask learning approach offers a robust solution for visual object tracking.