SIMULTANEOUS LOCALIZATION AND MAPPING OF UNMANNED ROBOTIC VEHICLE WITH APPLICATIONS TO AUTONOMOUS NAVIGATION IN A CHALLENGING ENVIRONMENT

Abstract

In this thesis we propose ORB-LINE-SLAM, a real-time hybrid point-line and only-line based visual Simultaneous Localizing and Mapping (SLAM) system for stereo cameras which can work in standard CPUs. This work is based on the ORB-SLAM3 open-source library. The motivation of the present thesis was to infuse line segments into the core of ORB-SLAM3 so as to overcome the challenges that it faces in demanding environments in which point features are not sufficient to estimate precisely the pose of the camera. The first novelty of the system is the introduction of an experimentally tuned adapting factor which aims to achieve a more efficient fusion of point and line features, as compared to classical methods which weigh the participation of points and lines in the optimization process equally. This necessity emerged from the complicated nature of the optimizations used in the line SLAM which assume that the endpoints of corresponding line segments are stable, although they are actually drifted throughout the infinite line that the line segment defines. In this way, we minimize the accumulating error in frames plentiful in point correspondences while we still take advantage of the rich information that line segments provide in low-textured environments in which the system determines insufficient orb matches. Moreover, to the best of our knowledge, this is the first open-source visual SLAM system that has the ability to effectively work by using line features exclusively. The biggest challenge related to only-line SLAM is that with the current existing line segment detectors and the restrictions that they set regarding the computational cost, the system can not define an adequate number of correspondences for the pose estimation. However, in our implementation by employing an efficient strategy for abstracting outliers and extracting lines from 2 scale levels with a sacrifice of computational cost, we significantly reinforce the efficacy of the system and partly overcome this impedance. Additional contributions of this thesis concern the expansion and adaptation of the components of ORB-SLAM3 to cope with line features, as well as the systematic comparison of different line feature detectors and error functions for the Bundle Adjustment (BA) process, to deduce which ones are the more effective for the visual SLAM problem. The results obtained from our experiments indicate that the proposed point-line method significantly improves the accuracy of ORB-SLAM3 in challenging conditions. Furthermore, the only-line method implemented in this work, despite its simplified nature, also manages to work in a standalone mode and to finish even difficult sequences with remarkable precision.