Transformer-augmented dual-branch siamese tracker with confidence-aware regression and adaptive template updating.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Siamese networks are effective for visual object tracking but struggle with static templates, limited context, and feature integration, especially during occlusion or clutter.
• Existing methods face challenges in maintaining tracking accuracy due to background variations and target appearance changes.

Purpose of the Study:

• To develop a robust and quality-aware visual object tracker that overcomes the limitations of traditional Siamese networks.
• To enhance tracking performance through transformer-based feature fusion and adaptive template updating.

Main Methods:

• Proposed TSDTrack, a transformer-augmented Siamese tracker utilizing a ResNet backbone for multi-scale feature extraction.
• Implemented a transformer module with global attention for enhanced semantic and spatial consistency.
• Introduced a confidence-aware branch (CAB) and regression distribution learning (RDL) for precise localization and quality assessment.
• Developed a confidence-gated template update strategy for adaptive appearance modeling.

Main Results:

• TSDTrack achieved state-of-the-art performance on benchmark datasets: LaSOT (55.5% success), GOT-10k (67.5% AO), OTB100 (71.6% AUC), and UAV123 (66.4% success).
• The tracker demonstrated superior accuracy and robustness compared to recent transformer-based and Siamese trackers.
• Confidence-aware and distribution learning branches improved localization precision under uncertainty.

Conclusions:

• TSDTrack offers a significant advancement in visual object tracking, particularly in challenging scenarios.
• The integration of transformers and quality-aware mechanisms provides a robust framework for adaptive tracking.
• The proposed methods effectively address limitations in context modeling and feature integration for Siamese trackers.