Synthetic Dataset for Quadcopter Detection Based on Frequency Propeller Signature

• Published in: 2023 IEEE Applied Imagery Pattern Recognition Workshop (AIPR) pp. 1 - 6
• Main Authors: Drouin, Marc-Antoine, Rainville, Marc-Andre, Picard, Michel, Stewart, Terrence C., Billy Djupkep Dizeu, Frank, Gagne, Guillaume
• Format: Conference Proceeding
• Language: English
• Published: IEEE 27.09.2023
• Subjects: Drones; Event Camera; labeled training dataset; Machine Learning (ML); neuromorphic camera; Neuromorphics; Propellers; Quadcopter; Quadrotors; Sensor systems; Synthetic data; Training; Uncrewed Aerial Systems (UAS)
• ISSN: 2332-5615
• DOI: 10.1109/AIPR60534.2023.10440658

The use of computer graphic tools typically associated with video games is a popular method to generate synthetic datasets for the training of machine learning algorithms. For optical detection of quadcopters, realistic imagery needs to be generated for multiple models of drones, in multiple types of environments and different flight profiles. By itself, the effort required to generate those virtual environments can be as important as flying actual drones. This is particularly true when a physics-based engine is required to model the behavior of the propellers. While some appearance-based drone detection methods may not need accurate propeller behavior, other detection methods that exploit the high-frequency and/or temporal signatures generated by rotating propellers do require such accurate simulations. This is especially the case for neuromorphic sensors, which are generally sensitive to the unique high-frequency visual signal from the propeller blades. We propose a synthetic approach to acquire training datasets for neuromorphic sensors using a flexible hardware system built from quadcopter components. This system allows efficient acquisition of training sets for drone detection sensors based on the propeller's temporal and/or frequency signature.