Anchor free based Siamese network tracker with transformer for RGB-T tracking.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Existing RGB-THERMAL tracking methods often rely on anchor-based algorithms and cross-correlation, limiting tracking success rates.
• Challenges include difficulties in improving target tracking accuracy and handling diverse environmental conditions.

Purpose of the Study:

• To propose a novel anchor-free tracking network, siamAFTS, for enhanced RGB-THERMAL target tracking.
• To address the limitations of current methods by improving the success rate of target tracking.

Main Methods:

• Developed the siamAFTS tracking network, an anchor-free approach utilizing a fully convolutional network with a Transformer module.
• Integrated channel and channel spatial attention modules to mitigate background interference.
• Eliminated anchor points to avoid hyperparameter tuning and reduce human intervention.

Main Results:

• The siamAFTS network demonstrated an improved success rate in target tracking across three benchmark datasets.
• The proposed model effectively reduces background interference, leading to more accurate bounding box predictions.
• Achieved superior performance compared to existing anchor-based trackers like MANET, DAPNet, SGT, and ADNet.

Conclusions:

• The anchor-free siamAFTS network offers a significant advancement in RGB-THERMAL target tracking.
• The integration of Transformer and attention modules enhances tracking robustness and accuracy.
• This approach simplifies the tracking process by removing anchor hyperparameter dependencies.