Siamese Regression Tracking With Reinforced Template Updating.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Siamese networks excel in visual tracking due to efficient localization.
• However, fixed templates and network parameters limit their ability to handle target appearance variations.
• Online parameter updates are often omitted for real-time performance.

Purpose of the Study:

• To propose a novel template updating method using reinforcement learning for Siamese regression trackers.
• To enhance the adaptability of Siamese trackers to dynamic changes in target appearance.
• To improve overall tracking performance by addressing limitations of fixed templates.

Main Methods:

• A template updating strategy is developed using an actor-critic framework.
• The actor network, trained via deep reinforcement learning, dynamically updates target templates each frame.
• The Siamese regression tracker's network parameters are also updated online.

Main Results:

• Experimental results on standard benchmarks demonstrate the effectiveness of both template and network updating.
• The proposed tracker, SiamRTU, shows significant improvements in handling appearance variations.
• Performance is competitive with state-of-the-art visual tracking approaches.

Conclusions:

• The proposed reinforcement learning-based template and network updating method effectively addresses appearance variations in Siamese trackers.
• SiamRTU offers a robust solution for visual tracking tasks requiring adaptation to dynamic environments.
• This approach enhances the practical applicability of Siamese networks in real-world scenarios.