Online Hierarchical Sparse Representation of Multifeature for Robust Object Tracking

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Sparse representation methods offer promising object tracking results but often neglect visual information correlation and spatial neighborhood details.
• Existing sparse coding techniques independently encode local image regions, missing broader spatial context crucial for robust tracking.

Purpose of the Study:

• To propose a robust object tracking algorithm that overcomes limitations of traditional sparse representation methods.
• To enhance tracking performance by incorporating spatial neighborhood information and complementary visual features.

Main Methods:

• Utilizing multiple complementary features for object appearance description, modeling both instantaneous and stable features.
• Implementing a two-stage sparse-coded method that considers spatial neighborhood information and computational efficiency for appearance reconstruction.
• Employing a particle filter framework to select the most reliable tracker, incorporating reliability measurement via transient and reconstructed appearance models.
• Incrementally updating the training set and template library based on current tracking outcomes.

Main Results:

• The proposed algorithm demonstrates superior tracking accuracy and robustness on challenging video sequences.
• Experimental results validate the effectiveness of integrating spatial information and complementary features in object tracking.
• The two-stage sparse-coded method balances representational power with computational feasibility.

Conclusions:

• The developed robust tracking algorithm effectively addresses the limitations of conventional sparse representation techniques.
• The integration of spatial neighborhood information and complementary features significantly improves tracking performance.
• The algorithm offers a reliable and accurate solution for object tracking in complex visual environments.