Robust RGB-T Tracking via Graph Attention-Based Bilinear Pooling.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Existing Convolutional Neural Network (CNN)-based RGB-T tracking methods struggle with partially informative target pairs.
• Global feature fusion in current methods is insufficient when target information is incomplete.

Purpose of the Study:

• To propose a novel four-stream oriented Siamese network (FS-Siamese) for improved RGB-T tracking.
• To address the limitations of existing methods in handling partially useful information for tracking.

Main Methods:

• Formulating multidomain multilayer feature map fusion as a multiple graph learning problem.
• Developing a graph attention-based bilinear pooling module for partial feature interaction exploration.
• Employing meta-learning to incorporate category information for enhanced efficiency and semantic representation.

Main Results:

• The proposed FS-Siamese network effectively explores partial feature interactions between RGB and thermal targets.
• Meta-learning enhances the efficiency and semantic consistency of feature embedding fusion.
• Experimental results on GTOT and RGBT234 datasets show superior performance compared to state-of-the-art methods.

Conclusions:

• The FS-Siamese network offers a promising solution for all-weather RGB-T tracking in intelligent transportation systems.
• The proposed graph attention-based fusion and meta-learning strategies significantly improve tracking accuracy and robustness.