Graph neural network-tracker: a graph neural network-based multi-sensor fusion framework for robust unmanned aerial...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Unmanned Aerial Vehicle (UAV) tracking is crucial for surveillance, security, and autonomous navigation.
• Existing tracking methods face challenges with accuracy and robustness, especially in complex scenarios.

Purpose of the Study:

• To propose a novel Graph Neural Network-based tracker (GNN-tracker) for Unmanned Aerial Vehicle (UAV) tracking.
• To enhance tracking accuracy, robustness, and identity consistency using graph-based spatial-temporal modeling and multi-sensor fusion.

Main Methods:

• Developed a GNN-tracker framework integrating graph-based spatial-temporal modeling and Transformer-based feature extraction.
• Employed multi-sensor fusion (optical, thermal) to improve tracking performance.
• Dynamically constructed spatiotemporal graphs for improved object association.

Main Results:

• GNN-tracker (fused) achieved 91.4% MOTA and 82.3% HOTA, outperforming state-of-the-art methods like TransT.
• Real-time performance demonstrated with high frames per second (58.9 FPS for fused data).
• Ablation studies confirmed the critical contribution of graph-based modeling and multi-sensor fusion.

Conclusions:

• GNN-tracker provides a highly accurate, robust, and efficient solution for UAV tracking.
• The framework effectively addresses real-world challenges in diverse conditions and sensor modalities.
• Graph-based modeling and multi-sensor fusion are essential components for superior UAV tracking performance.