This study presents an appearance based Simultaneous Localisation and Mapping (SLAM) algorithm that exploits the Graphics Processing Unit (GPU) as a general-purpose numerical coprocessor. This approach delivers immediate performance gains and improves algorithmic scalability.
The GPU has massive parallel computational resources that have traditionally been tied to rendering of graphics. Recent multidisciplinary research aimed at harnessing these resources and redirecting them for general-purpose computing has yielded performance increases on the order of 1.5 Ð 600x when compared to equivalent optimised CPU implementations. This research has been sparked in part by the release of the CUDA API, which provides an interface to the memory and arithmetic units of the GPU. The robotics community has ported common algorithms to the GPU, however previous studies on GPU accelerated SLAM have focussed only on a single or few components of the broader system. This study takes a holistic approach, accelerating all major components of SLAM.
The SLAM model used, based on Milford and WyethÕs RatSLAM, contains three major algorithmic components. The visual scene recognition component is intrinsically parallel. The place and direction recognition component is accelerated through novel use of GPU textures. The existing loop relaxation component is serial and is replaced by a minimally invasive probabilistic framework that uses a parallel numerical solver.
The final GPU accelerated implementation achieves an average speedup of 3x on modest laptop graphics hardware, when compared to a CPU implementation. The acceleration enables faster than real-time (FRT) offline mapping, and makes the computational demands of real-time large-scale mapping tractable.
Author:- Hilton Bristow
Source:-The University of Queensland