Justin Fletcher, Air Force Space Command; Ian McQuaid, Air Force Research Laboratory; Peter Thomas, Air Force Research Laboratory; Jeremiah Sanders, MD Anderson Cancer Center; Greg Martin, Centuari
Keywords: Computer Vision, Generative Adversarial Networks, Object Detection
Abstract:
This work introduces a convolutional neural network that detects geosynchronous Earth orbit resident space
objects in ground-based electro-optical telescope imagery. Model performance on a variety of object detection
tasks is analyzed, and an extension of the general object detection algorithm assessment framework relevant
to RSOs is described. Additionally, this work introduces two new datasets for the development and evaluation
of detection algorithms. We report a maximum F1 point of 0:971, corresponding to 0:973 precision and 0:969
recall at a localization threshold of 8 pixels, which is the highest reported performance on the SatNet dataset
known to the authors. Measured performance exceeds that of a classical detection algorithm, SExtractor,
evaluated on the same task. We demonstrate improved sensitivity to objects clusters with smaller apparent
separation by simulating low-frequency occurrences and augmenting natural training data. By incorporating
temporal information using a recurrent extension of a detection model, we further improve sensitivity to dim
and closely spaced objects.
Date of Conference: September 17-20, 2019
Track: Machine Learning for SSA Applications