Development of Attitude Sensor using Deep Learning (Beginner friendly paper for CV on a spacecraft)


(Ranet) #1

https://digitalcommons.usu.edu/cgi/viewcontent.cgi?article=4268&context=smallsat

Abstract

A new method for attitude determination utilizing color earth images taken with COTS visible light camera is
presented. The traditional earth camera has been used for coarse attitude determination by detecting the edge of the
earth, and therefore it can only provide coarse and 2-axis information. In contrast, our method recognizes the ground
pattern with an accuracy of sub-degrees and can provide 3-axis attitude information by comparing the detected ground
pattern and the global map. Moreover, this method has advantages in the size, mass and cost of the detector system
which consists of a cheap optical color camera and a single board computer. To demonstrate the method in space, we
have developed a sensor system named “Deep Learning Attitude Sensor (DLAS)”. DLAS uses COTS camera modules
and single board computers to reduce the cost. The obtained images are promptly analyzed with a newly developed
real-time image recognition algorithms.

A very fun, easy to read, paper, which talks about image segmentation/classification done live on a Raspberry Pi, attached to a probe in space, and the challenges they faced with computation cost.

The probe on which that component will be featured on: