LAUSR

Abstract CubeSats are limited in their physical size, which limits the physical size of payloads they can carry, thereby limiting the image quality of CubeSat Earth Observation missions. Algorithms exist that combine partially overlapping images to produce better output image quality. These algorithms may either improve the signal-to-noise ratio via averaging, increase resolution via super-resolution or merely remove redundant information via mosaicing. Such algorithms typically only function properly if the geometric transformations between consecutive images are known with high accuracy. These algorithms can either be applied terrestrially or on-board a satellite. Downloading large raw image data sets for terrestrial processing is impractical for a CubeSat mission, and therefore an on-board solution is desirable. This paper discusses how to accurately determine the transformation between consecutive images on-board, laying the foundation for efficient on-board de-noising, super-resolution and mosaicing. Two common methods used to determine translation – normalised cross correlation (NCC) and phase correlation – are investigated. From simulated results, NCC is shown to be the better candidate for our application. NCC achieves sub-pixel accuracy by making use of polynomial least squares regression. NCC is well suited for implementation on a satellite platform where images are captured in quick succession, resulting in partially overlapping images with little rotation between frames. We compare two potential hardware platforms – the MicroZed 7020 and Jetson TK1 – and then describe how we implemented our proposed solution onto the former, using a hardware description language. The effect of tuning NCC's parameters – specifically window and template size – is investigated with regards to hardware utilisation and accuracy. Software simulation and firmware-implementation results, using simulated data, are compared and discussed. We conclude by discussing expected full system runtime and energy efficiency.