AD-VAT+: An Asymmetric Dueling Mechanism for Learning and Understanding Visual Active Tracking

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Visual Active Tracking (VAT) involves autonomously controlling a tracker's motion based on visual input to follow a target.
• Developing robust trackers for VAT is challenging, especially in dynamic environments with occlusions or complex movements.

Purpose of the Study:

• To propose a novel adversarial reinforcement learning (RL) method, AD-VAT, for robust Visual Active Tracking (VAT).
• To enhance the method further with environment augmentation and two-stage training (AD-VAT+) for improved performance in challenging scenarios.

Main Methods:

• Introduced an Asymmetric Dueling mechanism where the tracker and target act as opposing agents.
• The target agent receives additional information (tracker's observations and actions) and predicts the tracker's reward as an auxiliary task.
• Employed advanced environment augmentation and two-stage training strategies for the AD-VAT+ variant.

Main Results:

• The asymmetric dueling mechanism strengthens the target, consequently inducing a more robust tracker.
• AD-VAT+ demonstrated faster training convergence compared to baseline methods.
• Experimental results in 2D and 3D environments showed more robust tracking behaviors in diverse testing scenarios.

Conclusions:

• The proposed AD-VAT and AD-VAT+ methods significantly improve the robustness and efficiency of Visual Active Tracking.
• The asymmetric dueling mechanism is effective in creating stronger opponents, leading to better tracker learning.
• The approach shows potential for real-world applications, as evidenced by demonstrations on real-world videos.