Mobile-URSONet: an Embeddable Neural Network for Onboard Spacecraft Pose Estimation

TL;DR

This work tackles onboard spacecraft pose estimation from monocular images under tight compute constraints. It adapts URSONet into Mobile-URSONet by swapping in a MobileNet-v2 backbone and optimizing bottleneck size and orientation encoding, notably using soft classification to improve orientation accuracy. The approach achieves up to $178$× fewer parameters with accuracy degradation not exceeding roughly $4$× compared to URSONet, and orientation accuracy improves from about $32^{\circ}$ (regression) to around $6.3^{\circ}$ (best soft-classification). The model demonstrates favorable accuracy-density and generalization potential on the SPEED dataset, and the authors point to quantization, pruning, and MPSoC deployment as future steps to reinforce onboard applicability and real-time performance.

Abstract

Spacecraft pose estimation is an essential computer vision application that can improve the autonomy of in-orbit operations. An ESA/Stanford competition brought out solutions that seem hardly compatible with the constraints imposed on spacecraft onboard computers. URSONet is among the best in the competition for its generalization capabilities but at the cost of a tremendous number of parameters and high computational complexity. In this paper, we propose Mobile-URSONet: a spacecraft pose estimation convolutional neural network with 178 times fewer parameters while degrading accuracy by no more than four times compared to URSONet.

Figures (2)

• Figure 1: Mobile-URSONet architecture
• Figure 2: Position and orientation error by distance for our 12-bins model