Adaptive low-rank subspace learning with online optimization for robust visual tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Sparse and low-rank models are common for visual tracking appearance subspace learning.
• Existing methods often focus only on sparsity or low-rankness of coefficients, proving insufficient for complex sequences.
• Incremental optimization for nuclear and column sparse norms is challenging with sequential data.

Purpose of the Study:

• To develop a novel Low-Rank Subspace Learning with Adaptive Penalization (LSAP) framework for robust visual tracking.
• To simultaneously learn subspace basis, low-rank coefficients, and column sparse errors for appearance subspace formulation.
• To introduce adaptive penalization for improved robustness on corrupted datasets.

Main Methods:

• The LSAP framework simultaneously learns subspace basis, low-rank coefficients, and column sparse errors.
• A Hadamard product-based regularization is introduced for adaptive penalization of coefficients.
• An efficient incremental optimization scheme is developed for online visual tracking.

Main Results:

• Adaptive penalization significantly improves LSAP robustness on severely corrupted datasets.
• The proposed tracker demonstrates superior performance compared to state-of-the-art methods.
• Experiments conducted on 50 challenging video sequences validate the effectiveness of LSAP.

Conclusions:

• LSAP offers a robust approach to visual tracking by effectively formulating appearance subspace.
• The adaptive penalization and efficient optimization scheme contribute to enhanced tracking performance.
• The proposed method advances the field of subspace-based visual tracking.