Thermal Infrared Tracking Method Based on Efficient Global Information Perception.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Current Thermal InfraRed (TIR) tracking methods struggle with occlusion and similar target interference.
• Efficient global information perception is crucial for robust TIR object tracking.

Purpose of the Study:

• To develop a novel TIR tracking method that overcomes limitations in occlusion and interference.
• To enhance the efficiency and accuracy of TIR object tracking through advanced feature extraction and fusion.

Main Methods:

• Utilized a Transformer structure with Focal Transformer for efficient global semantic information extraction.
• Incorporated relative position encoding in the Transformer for improved feature fusion and positional awareness.
• Employed a heterogeneous bi-prediction head (fully connected for classification, convolutional for regression) for object prediction.
• Generated the LaSOT-TIR dataset using a generative adversarial network to address data scarcity.

Main Results:

• The proposed method demonstrated superior performance compared to state-of-the-art trackers on benchmark TIR datasets (VOT2015-TIR, VOT2017-TIR, PTB-TIR, LSOTB-TIR).
• Achieved outstanding results particularly in scenarios with severe occlusion and interference from similar objects.
• The Transformer-based approach effectively models semantic information and improves tracking accuracy and robustness.

Conclusions:

• The developed TIR tracking method effectively addresses challenges of occlusion and similar target interference.
• The integration of Transformer architecture with global information perception significantly advances TIR tracking capabilities.
• The method offers a robust and accurate solution for real-world TIR tracking applications.