Improvements to SLAM architecture on embedded systems

Abstract

This paper presents the development of various SLAM (Simultaneous Localization and Mapping) techniques. The focus will be on visual based SLAM techniques. Visual SLAM is where one or more cameras are involved in the process of mapping and localization. Visual SLAM’s accuracy improves as the knowledge of both the map and robot path are improved over time due to more data being acquired and landmarks being observed from different positions in the environment. However, this location can only be taken from a pre-existing map of the environment it was in. There are a lot of factors that need to be considered when collecting the results from SLAM, factors such as computational limits due to hardware, accuracy, noise, and speed of the algorithm may impact SLAM's results. Numerous algorithms present such as ORB-SLAM2 (Oriented FAST and Rotated BRIEF SLAM) [1] are not optimized to fully maximize the available hardware. ORBSLAM2 however is a very well-rounded algorithm and works across different hardware architectures. Optimizations where the GPU (Graphical Processing Unit) can be used will improve performance in the algorithm. The objective would be to further improve the SLAM algorithm by taking advantage of hardware provided. The goals would be to add extra functionalities and increase performance on embedded systems.