RGBT Tracking via Noise-Robust Cross-Modal Ranking.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Existing Red-Green-Blue-Depth (RGBT) tracking algorithms often struggle with background clutter, leading to inaccurate target localization within bounding boxes.
• This limitation impacts the reliability of RGBT tracking in complex, real-world scenarios.

Purpose of the Study:

• To develop a novel algorithm, noise-robust cross-modal ranking, to suppress background effects in RGBT tracking.
• To improve the robustness and accuracy of RGBT tracking by addressing noise interference in cross-modal fusion and seed labels.

Main Methods:

• Proposed a soft cross-modality consistency mechanism to balance collaboration and heterogeneity during modality fusion.
• Introduced optimal seed learning to mitigate label noise arising from irregular object shapes and occlusions.
• Implemented individual feature ranking and cross-feature consistency to leverage modality complementarity and structural information.
• Developed a unified optimization framework for efficient model convergence.

Main Results:

• The proposed noise-robust cross-modal ranking algorithm effectively suppresses background effects in RGBT tracking.
• Experiments on GTOT and RGBT234 datasets show superior performance compared to state-of-the-art tracking methods.
• The approach demonstrates both effectiveness and efficiency in challenging tracking scenarios.

Conclusions:

• The novel algorithm significantly enhances RGBT tracking accuracy and robustness by effectively handling background clutter and noise.
• The proposed methods for cross-modal fusion and seed learning offer a promising direction for future RGBT tracking research.
• The developed unified optimization framework provides an efficient solution for practical RGBT tracking applications.