Toward an optimized 3D Monocular SLAM

Simultaneous Localization and Mapping (SLAM) is a fundamental problem for an autonomous mobile robot. This topic has gained momentum in UCL's robotics labs after UCL's autonomous drone project has been initialized. During the last years, the main objective of this project was to build an autonomous drone using a low-cost quadcopter platform. One way to undertake this problem was by implementing the Monocular SLAM. However, the system was not efficient and powerful. This is why the main goal of this master's thesis is to search and find new techniques to allow better performance of the UCL's Monocular SLAM. Therefore, we will review the different elements constituting the SLAM. We will investigate the cause of the issues and propose some solutions to address those problems. Different algorithms and optimizations have been implemented in order to improve the performance of the Monocular SLAM such as a new feature detector & descriptor "ORB", an improvement in the bundle local adjustment and also new techniques to remove landmarks. After testing the new UCL's SLAM’s version, the results show an improvement in the performance and efficiency of the monocular SLAM. The algorithm now able to create accurate maps without slowing down the hardware.