Learning Collaborative Sparse Representation for Grayscale-Thermal Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Object tracking often relies on single-source data, limiting performance in varied conditions.
• Integrating complementary features like grayscale and thermal data can enhance tracking robustness.

Purpose of the Study:

• To develop a robust object tracking algorithm by adaptively fusing grayscale and thermal video information.
• To introduce a novel collaborative tracking approach for challenging scenarios.

Main Methods:

• Proposed an adaptive fusion scheme using collaborative sparse representation within a Bayesian filtering framework.
• Jointly optimized sparse codes and modality weights in an online manner.
• Created a diverse benchmark dataset of 50 grayscale-thermal video sequences with consistent annotations.

Main Results:

• The proposed algorithm demonstrated superior performance compared to state-of-the-art trackers on grayscale and grayscale-thermal inputs.
• Extensive experiments validated the effectiveness of the adaptive fusion approach.

Conclusions:

• Adaptive fusion of grayscale and thermal data significantly improves object tracking robustness.
• The developed benchmark and algorithm provide a foundation for future research in multi-modal tracking.