Spacecraft pose estimation, on the edge

Monocular pose estimation is the challenge of estimating the relative attitude and position of a target using a single camera. The computer vision community has demonstrated high accuracy and robust algorithms for pose estimation in terrestrial applications. However the space environment presents unique challenges. Space-grade hardware for example greatly constrains the power consumption and memory bandwidth of such algorithms. As such, the computer science trend of larger and larger networks is not feasible in a deployed hardware application.

Master’s project (can be adapted for TP-IV)

Objective

The aim of this project is to demonstrate spacecraft pose estimation algorithms running on small edge devices.

Tasks

Survey available pose estimation algorithms; select two contenders.

Deploy the algorithms to an edge device (currently a JETSON NANO is available).

Compare performance of the two algorithms.

  • latency
  • memory
  • accuracy

Prerequisites

  • Proficiency in python
  • Familiarity with pytorch
  • Understanding of the camera model (world frame to image plane linear algebra)
  • Experience with micro-controllers or edge-devices a plus

References

A Survey on Deep Learning-Based Monocular Spacecraft Pose Estimation: Current State, Limitations and Prospects

https://arxiv.org/abs/2305.07348
Acta Astronautica 2023

Wide-Depth-Range 6D Object Pose Estimation in Space

https://arxiv.org/abs/2104.00337

CVPR 2021

DVMNet: Computing Relative Pose for Unseen Objects Beyond Hypotheses
https://arxiv.org/pdf/2403.13683
CVPR 2024

Contact

This project will be conducted in collaboration with CVLab. Contact [email protected] for more information.