AMST²: aggregated multi-level spatial and temporal context-based transformer for robust aerial tracking.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Existing visual trackers often use spatial or temporal information separately.
• This limits their ability to leverage complementary advantages for improved tracking.

Purpose of the Study:

• To develop a novel transformer-based visual tracker integrating multi-level spatial and temporal contexts.
• To enhance robustness in complex aerial tracking scenarios.

Main Methods:

• Proposed a transformer-based model incorporating multi-level spatial and temporal context information.
• Introduced an aggregation encoder to integrate refined similarity maps.
• Utilized a lightweight network backbone for efficient tracking.
• Implemented a feature update mechanism retaining initial template information for robustness.

Main Results:

• The proposed tracker effectively integrates global spatial and temporal contexts.
• Achieved superior performance in complex aerial scenarios, addressing occlusion, blur, and scale variations.
• Demonstrated state-of-the-art results on seven aerial tracking benchmarks.
• Outperformed existing methods in both real-time processing speed and tracking accuracy.

Conclusions:

• The novel approach offers a robust and efficient solution for aerial visual tracking.
• The integrated multi-level spatial and temporal features are crucial for handling challenging aerial environments.
• The tracker provides a significant advancement in real-time object tracking performance.