Researchers led by Charlotte Sieh have built Pelorus, a lidar-inertial odometry method that works out how a robot, car or drone moves while keeping its memory use bounded, according to a paper on arXiv. High-resolution lidars produce millions of points a second, which small onboard computers have to process as they arrive.
Pelorus keeps a map of fixed size at several resolutions, drops its least relevant parts when it is full, and gives less weight to observations that do not fit. It runs in parallel on processors and graphics chips.
On five sequences from three public datasets, it had the lowest relative pose error in seven of ten measures with a small memory footprint, and with work moved to the graphics chip it put the lowest load on the processor of the methods compared. On an NVIDIA Jetson Orin NX it handles a 128-beam lidar at 10 Hz. Pelorus is open source.



