Robust Visual SLAM in Challenging Environments with Low-texture and Dynamic Illumination

Abstract

In the last years, visual Simultaneous Localization and Mapping (SLAM) has played a role of capital importance in rapid technological advances, e.g. mo- bile robotics and applications such as virtual, augmented, or mixed reality (VR/AR/MR), as a vital part of their processing pipelines. As its name indicates, it comprises the estimation of the state of a robot (typically the pose) while, simultaneously, incrementally building and refining a consistent representation of the environment, i.e. the so-called map, based on the equipped sensors.

Despite the maturity reached by state-of-art visual SLAM techniques in controlled environments, there are still many open challenges to address be- fore reaching a SLAM system robust to long-term operations in uncontrolled scenarios, where classical assumptions, such as static environments, do not hold anymore. This thesis contributes to improve robustness of visual SLAM in harsh or difficult environments, in particular:

- Low-textured Environments, where traditional approaches suffer from an accuracy impoverishment and, occasionally, the absolute failure of the system. Fortunately, many of such low-textured environments contain planar elements that are rich in linear shapes, so an alternative feature choice such as line segments would exploit information from structured parts of the scene. This set of contributions exploits both type of features, i.e. points and line segments, to produce visual odometry and SLAM algorithms robust in a broader variety of environments, hence leveraging them at all instances of the related processes: monocular depth estimation, visual odometry, keyframe selection, bundle adjustment, loop closing, etc. Additionally, an open-source C++ implementation of the proposed algorithms has been released along with the published articles and some extra multimedia material for the benefit of the community.