The detection of meteors, a luminous phenomenon resulting from the entry of extraterrestrial material into the atmosphere, is a topic of interest for astronomers. The Meteorix project aims to perform this detection from a nanosatellite, a CubeSat 3U, in order to overcome the constraints associated with ground detection. This thesis is part of this project and focuses on enriching and optimizing a space detection processing chain while adhering to two significant constraints: real-time processing (40 ms per image) on a low-power system-on-chip (SoC) limited to 10 W. This work addresses key challenges related to optimizing algorithms to ensure real-time processing and identifying possible approximations that do not compromise the detection quality.
Spatial detection has to take into account the movements of the satellite and the observed scene, rendering commonly used algorithms ineffective. For this reason, the Meteorix project application relies on an optical flow algorithm (Horn & Schunck) to estimate the various apparent motions within the scene. The complete application achieves a detection rate of 96%.
To meet execution time and energy constraints, the optical flow algorithm was parallelized (using OpenMP and SIMD) and then optimized through algorithmic transformations. These include modifications to the code semantics, two iteration pipeline schedulings, and a multi-scale approximate estimation. The optimizations achieved allow for speedups ranging from x15 to x28 while reducing energy consumption by a factor of three compared to the compiler-optimized version. Tests on embedded systems, such as the Jetson Orin Nano and the Orange Pi 5+, identified configurations that respect the 10 W energy limit while maintaining a detection rate of 96%.