Tracking by Joint Local and Global Search: A Target-Aware Attention-Based Approach

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Tracking-by-detection is a common framework for single-object tracking.
• Local search mechanisms in tracking-by-detection are vulnerable to occlusion and fast motion.

Purpose of the Study:

• To propose a novel target-aware attention mechanism (TANet) for robust single-object tracking.
• To integrate TANet into a tracking-by-detection framework for joint local and global search.

Main Methods:

• Feature extraction from target object patches and video frames.
• A decoder network generates target-aware global attention maps.
• Adversarial training with appearance and motion discriminators enhances attention prediction.

Main Results:

• The proposed TANet improves tracking performance in challenging scenarios.
• Experiments on benchmark datasets demonstrate the effectiveness of the algorithm.
• The method achieves robust tracking through integrated local and global search.

Conclusions:

• TANet offers a general and effective solution for robust single-object tracking.
• The joint local and global search strategy enhances tracking accuracy and resilience.
• Adversarial training further refines attention prediction for improved spatial-temporal consistency.