Fast processing of grid maps using graphical multiprocessors

Abstract

Grid mapping is a very common technique used in mobile robotics to build a continuous 2D representation of the environment useful for navigation purposes. Although its computation is quite simple and fast, this algorithm uses the hypothesis of a known robot pose. In practice, this can require the re-computation of the map when the estimated robot poses change, as when a loop closure is detected. This paper presents a parallelization of a reference implementation of the grid mapping algorithm, which is suitable to be fully run on a graphics card showing huge processing speedups (up to 50×) while fully releasing the main processor, which can be very useful for many Simultaneous Localization and Mapping algorithms.