A Unified GNN-CV Framework for Intelligent Aerial Situational Awareness.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Aerospace Engineering

Background:

• Aerial situational awareness (SA) is complex, involving dynamic entities and spatio-temporal relationships.
• Current deep learning methods for SA often lack a holistic, vision-centric approach vital for human decision-making.
• Existing models struggle with diverse data modalities and integrating them for comprehensive SA.

Purpose of the Study:

• To develop a unified Graph Neural Network-Computer Vision (GNN-CV) framework for operational-level aerial SA.
• To bridge the gap between deep learning capabilities and the human-centric perspective needed for effective decision-making.
• To process radar-map-like representations using mature computer vision architectures for diverse SA tasks.

Main Methods:

• Proposed a unified GNN-CV framework integrating sparse entity attribute transformation graph neural networks (SET-GNNs).
• Developed methods for large-scale radar map reconstruction and integrated feature extraction.
• Employed specialized two-stage pre-training and adaptable downstream task networks for aerial SA.

Main Results:

• Achieved end-to-end recognition accuracy exceeding 90.1% for aerial swarm partitioning and configuration recognition.
• Configuration recognition accuracy surpassed 85.0% in tactical scenarios with varied flight intervals and formation types.
• Maintained accuracy above 80.4% even with significant position/heading disturbances, operating within millisecond response cycles.

Conclusions:

• Leveraging mature computer vision techniques significantly enhances the efficacy, resilience, and coherence of intelligent SA.
• The proposed GNN-CV framework offers a robust solution for complex operational-level aerial SA tasks.
• The framework demonstrates superior performance in recognizing aerial formations under challenging conditions.