Siddharth Dave, York University; Ryan Clark, York University; Gabriel Chianelli, York University; Regina Lee, York University
Keywords: Machine learning; star tracker camera; space situational awareness; resident space object; image processing
Abstract:
The growing number of resident space objects (RSOs) in low Earth orbit (LEO) presents a concern for the future of our planets space industry. The US space surveillance network (SSN) currently detects, tracks and catalogues upwards of 20,000 RSOs using a suite of dedicated ground and space-based sensors. A potential optical source of dispersed in-orbit space data for space situational awareness (SSA) are star trackers, a commonly found attitude determination instrument on satellites. This paper discusses the results of a novel recurrent convolutional neural network (R-CNN) based algorithm designed to make position and velocity estimates of in-orbit RSOs captured by a commercial-grade star tracker. Due to the low-resolution quality of star tracker images, the designed algorithm takes advantage of the high frame rate and short exposure time to estimate RSO state vector. Furthermore, by characterizing the streak patterns obtained from image sequences, followed by a deconvolution of star tracker attitude changes and orbital motion, relative position and velocity estimates can be achieved. The R-CNN design initially trains on synthetically generated noiseless images and is validated by images generated using an analytical image simulator. The objective of this research is to discuss how the designed algorithm can facilitate RSO orbit estimation for SSA using low-resolution star tracker. Preliminary results indicate that sub-pixel estimation and integration-time streaks of RSOs from star tracker images are viable candidates for state vector estimation.
Date of Conference: September 15-18, 2020
Track: Machine Learning Applications of SSA