Comparison of modern open-source Visual SLAM approaches

Simultaneous localization and mapping (SLAM) is one of the fundamental areas of research in robotics and environment reconstruction. State-of-the-art solutions have advanced significantly in terms of mapping quality, localization accuracy and robustness. It becomes possible due to modern stable solvers in the back-end, efficient outlier rejection techniques and diversified front-end: unique features, topologically segmented landmarks, and high-quality sensors. Among the variety of open-source solutions, several promising approaches provide results which are difficult to be reproduced on standard datasets, especially if there is no description for dataset adaptation. The goal of the article is to figure out, which techniques of robots’ localization are the most promising for further use in related disciplines for engineers and researchers. The main contribution is a comparative analysis of state-of-the-art open-source Visual SLAM methods in terms of localization precision for versatile environments. The algorithms are assessed based on accuracy, computational performance, robustness and fault tolerance. Additionally, the survey and comparison of the datasets used for methods evaluation are provided as well as practical recommendations of usage scenarios for further research.