A Transformer-based Approach to Near-Earth Asteroid Detection

This paper presents a novel approach to Near-Earth Asteroid (NEA) detection, founded upon deep learning principles. The proposed learning architecture is a pre-trained video transformer network, which is fine-tuned on synthetic data. The video transformer learns the spatio-temporal relationships between frames of a tracklet of exposures, and determines if an asteroid is present within the tracklet. The synthetic tracklet generation algorithm is detailed, in which the known NEA population statistics are utilized as the basis for data generation. The trained network is verified with real tracklets from the Siding Spring Observatory.