The Research Of Parallel FastSLAM Algorithm Based On CUDA

The key to autonomous of mobile robot navigation is Simultaneous Localization and Mapping(SLAM); It is an urgent problem to make the SLAM process more efficient.Fast SLAM is based on the Rao-Blackwellized theorem which decomposes the SLAM problem into a robot path estimation and landmark estimation process. It is good to reduce the computational complexity of the algorithm and the running time. Nevertheless, there still be a lot of research focused on how to improve the performance of SLAM. With the rapid development of graphics card technology, NVIDIA launched a general-purpose parallel computing architecture CUDA in 2007. The GPU can be used to solve complex computational problems and data parallel problems under this architecture. To achieve the purpose of accelerating, developers can use the CUDA C language to write the program.Fast SLAM algorithm is the basis of the thesis. The resampling process and estimation process are the points we considered. We improved the algorithm by using CUDA, which make full use of CPU and GPU, so as to achieve the purpose of accelerating. The main job of this paper includes:(1) According to resampling process of the Fast SLAM, we chose system resampling algorithms, and redesign the sampling rule. The left boundary is l =(c k-u)?N +2 and the right() 1ir =c k +w k-u ?N +, which make the data independent and satisfy the data parallelism. We use of "extern C" as a interface between CUDA program and general C++ program;(2) The extend Kalman filter is used to landmark estimation of Fast SLAM, which has large mount of matrix calculation. We use CUDA C to optimized the computing of EKF.