Implementação paralela de algoritmos para localização e mapeamento simultâneos com uma única câmera

Abstract

This work is on the Simultaneous Localization and Mapping (SLAM) problem for the special case where a single camera is used, called Monocular SLAM. The main goal of this work is the investigation of solutions that would be capable to efficiently tackle the Monocular SLAM problem, exploring parallelism and its implementation on Graphics Processing Units (GPU). A new parallel algorithm for pose estimation based on particle filters and the Rao-Blackwell factorization is proposed and implemented. As in FastSLAM, a new map is estimated for each particle and the proposed distribution for the next pose provides information derived from the last observation. However, unlike FastSLAM, our algorithm incorporates this information using a second sub-particle filter for each particle. This enables the adaptation of the algorithm to a wide range of locomotion models. Parallel versions of the proposed algorithm using a SICK laser and of the Monte Carlo algorithm have also been implemented. In the former Speedups of about 500 times were obtained in comparison with the sequential version, and the later has shown to be able to localize the mobile robot in real time using over 100,000 particles with measured performance over 100 times when compared to the sequential implementation of the method.