A Siamese tracker with "dynamic-static" dual-template fusion and dynamic template adaptive update.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Siamese network-based visual trackers offer a balance of speed and accuracy.
• Static templates struggle with target appearance changes, while dynamic templates risk noise contamination.

Purpose of the Study:

• To propose a Siamese tracker with a "dynamic-static" dual-template fusion and adaptive dynamic template update.
• To improve robustness and accuracy in visual object tracking under challenging conditions.

Main Methods:

• A novel tracker combining static and real-time updated dynamic templates.
• Utilizing an adaptive update strategy for the dynamic template to handle appearance variations and suppress noise.
• Building upon DaSiamRPN and UpdateNet architectures.

Main Results:

• On the VOT2016 dataset, robustness increased by 23% and Expected Average Overlap (EAO) by 9.0% compared to the baseline.
• On the OTB100 dataset, precision improved by 0.8% and success rate by 0.4%.

Conclusions:

• The proposed dual-template fusion and adaptive update strategy significantly enhances visual tracking performance.
• The method achieves comprehensive real-time tracking performance on benchmark datasets.