Inverse sparse tracker with a locally weighted distance metric

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Sparse representation is a key technique for enhancing visual tracking accuracy.
• Existing methods often require solving multiple optimization problems per frame, limiting efficiency.
• Standard distance metrics are sensitive to occlusions and illumination variations.

Purpose of the Study:

• To propose an efficient and robust sparsity-based visual tracking algorithm.
• To introduce an inverse sparse representation formulation for computational efficiency.
• To develop a locally weighted distance metric for improved robustness against appearance changes.

Main Methods:

• An inverse sparse representation formulation reconstructs target templates with particles, enabling single l1 optimization for particle weighting.
• A locally weighted distance metric is designed, mathematically analyzed, and implemented using temporal and spatial continuity for weight assignment.
• The approach explicitly considers appearance changes due to occlusion and shape deformation.

Main Results:

• The proposed algorithm demonstrates significant efficiency by solving only one optimization problem.
• Experimental validation on 15 challenging sequences shows superior performance compared to state-of-the-art methods.
• The locally weighted distance metric effectively handles partial occlusion and illumination changes.

Conclusions:

• The developed sparsity-based tracker offers improved efficiency and robustness.
• The inverse sparse representation and locally weighted distance metric are effective components for visual tracking.
• The algorithm shows strong potential for real-world applications requiring accurate and efficient visual tracking.