Isolation and predictability of embeddable Neural Networks for Onboard Spacecraft

Debris removal is crucial for the future of low earth orbit operations as the exploitation of this orbit grows. A significant increase of debris will accompany these new constellations. Estimating the relative position and orientation (commonly called Pose estimation) of a known but uncooperative space-craft from a monocular image is an essential computer vision application that allows improving the autonomy of in-orbit spacecraft operations. In an ongoing Mitacs project, set to complete in August 2022, we proposed the Mobile-URSONet, an embeddable Deep Neural Network (DNN) architecture that can be used for spacecraft pose estimation and that is compatible with the constraints of onboard spacecraft computers. The main objective of this proposed Mitacs project is to go a step further in the development of an embeddable Deep Neural Network (DNN) accelerator, by building a prototype that demonstrates the proper functioning of a system-level design flow that explores and supports implementing various tradeoffs for a pose estimation application based the Mobile-URSONet IPs.

Faculty Supervisor:

Yvon Savaria;Guy Bois

Student:

Partner:

Space Codesign Systems

Discipline:

Engineering

Sector:

Information and cultural industries

University:

Polytechnique Montréal

Program: