DP-Siam: Dynamic Policy Siamese Network for Robust Object Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Object tracking faces a trade-off between real-time performance and accuracy.
• Existing methods struggle to balance speed and precision in dynamic environments.

Purpose of the Study:

• To propose a novel dynamic policy gradient Agent-Environment architecture with Siamese network (DP-Siam) for improved object tracking.
• To enhance both accuracy and expected average overlap in real-time tracking applications.

Main Methods:

• DP-Siam utilizes reinforcement learning for offline training to predict optimal object locations.
• The architecture comprises an Agent network (state prediction), Environment network (Q-value estimation), and Siamese network (heatmap generation).
• During online tracking, the Environment network verifies the Agent network's actions.

Main Results:

• Extensive experiments were conducted on six benchmark datasets (OTB2013, OTB50, OTB100, VOT2015, VOT2016, VOT2018).
• DP-Siam demonstrated significant performance improvements over current state-of-the-art trackers.
• The proposed method effectively balances real-time performance with high tracking accuracy.

Conclusions:

• DP-Siam offers a robust solution to the real-time vs. accuracy challenge in object tracking.
• The novel architecture and training methodology lead to superior tracking performance.
• This approach sets a new benchmark for object tracking systems.