Beyond Greedy Search: Tracking by Multi-Agent Reinforcement Learning-Based Beam Search.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Current visual trackers often use greedy search, which can fail in challenging scenarios like occlusion and fast motion.
• Tracker drift can lead to accumulated errors and unreliable future predictions.

Purpose of the Study:

• To propose a novel visual tracking strategy that overcomes limitations of greedy search.
• To improve tracking accuracy and robustness in difficult conditions by reducing accumulated errors.

Main Methods:

• Introduced BeamTracking, a multi-agent reinforcement learning-based beam search tracking strategy.
• Formulated visual tracking as a sample selection problem using parallel decision-making processes.
• Employed bi-GRU for state representation and multiple agents for diverse trajectory prediction.

Main Results:

• The proposed BeamTracking strategy effectively identifies trajectories with fewer accumulated errors.
• Extensive experiments on seven benchmark datasets demonstrated the algorithm's effectiveness.
• Validated improved performance compared to traditional greedy search methods.

Conclusions:

• BeamTracking offers a more robust and accurate approach to visual object tracking.
• The multi-agent reinforcement learning and beam search combination is effective for handling challenging tracking scenarios.
• This method provides a promising direction for future visual tracking research.