A Standardized Maneuver Pattern Library and Dual-View Framework for Multi-View Maneuver Classification.

Area of Science:

• Aerospace Engineering
• Artificial Intelligence
• Computer Vision

Background:

• Aerial vehicle maneuver pattern classification is crucial for airspace management.
• Existing methods struggle with limited, imbalanced, and noisy real-world flight data, hindering generalization.
• Lack of diverse, standardized datasets limits robust aerial maneuver recognition.

Purpose of the Study:

• To develop a robust aerial maneuver pattern classification method overcoming data limitations.
• To create a standardized dataset for training and evaluating maneuver classification models.
• To propose a lightweight deep learning model capable of effective sim-to-real transfer.

Main Methods:

• Constructed a Maneuver Pattern Library using X-Plane 12, featuring five maneuver primitives with noise reduction and interpolation.
• Developed DualView-LiteNet, a Siamese convolutional network processing dual-view trajectory data for spatial and temporal feature learning.
• Evaluated the model on simulated data and a real-world ADS-B dataset without fine-tuning.

Main Results:

• DualView-LiteNet achieved 97.64% accuracy on the benchmark dataset, with precision, recall, and F1-scores reaching 0.98.
• The model demonstrated strong generalization and sim-to-real transfer capability on real-world ADS-B data.
• The proposed approach offers improved robustness and effectiveness compared to baseline models.

Conclusions:

• The developed dataset and DualView-LiteNet provide an effective solution for aerial maneuver pattern classification.
• The sim-to-real transfer capability highlights the practical feasibility and robustness of the proposed method.
• This work establishes a reliable reference for future research in aerial vehicle behavior analysis.