Hierarchical Spatiotemporal Graph Regularized Discriminative Correlation Filter for Visual Object Tracking.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Visual object tracking is crucial for advanced vision and robotics.
• Discriminative correlation filters (DCFs) offer speed and accuracy but degrade due to boundary effects and lack of geometric constraints.

Purpose of the Study:

• To propose a novel method for robust object tracking using hierarchical spatiotemporal graph-regularized correlation filters.
• To address the limitations of existing DCF methods by incorporating spatial and temporal structural information.

Main Methods:

• Decomposing target samples into deep channels to construct spatial graphs.
• Compressing temporal information using principal component analysis to build temporal graphs.
• Constraining learned correlation filters as eigenvectors of spatiotemporal graph Laplacians within a novel objective function.

Main Results:

• The proposed algorithm demonstrates excellent performance compared to 33 state-of-the-art trackers.
• Evaluated on six challenging benchmark datasets, showing significant improvements in robust object tracking.

Conclusions:

• Hierarchical spatiotemporal graph regularization offers a robust approach to visual object tracking.
• The novel objective function and solution method effectively integrate spatiotemporal constraints into DCFs, outperforming existing methods.