The gold standard method for tridimensional reconstruction and camera localization from a set of images is well known to be Bundle Adjustment (BA). Although BA was regarded for years as a costly method restricted to the offline domain, several real time algorithms based on BA flourished in the last decade. However those algorithms were limited to perform SLAM in small scenes or only Visual Odometry. In this work we built on excellent algorithms of the last years to design from scratch a Monocular SLAM system that operates in real time, in small and large, indoor and outdoor environments, with the capability of wide baseline loop closing and relocalization, and including full automatic initialization. Our survival of the fittest approach to select the points and keyframes of the reconstruction generates a compact and trackable map that only grows if the scene content changes, enhancing lifelong operation. We present an exhaustive evaluation in 27 sequences from the most popular datasets achieving unprecedented performance with a typical localization accuracy from 0.2% to 1% of the trajectory dimension in scenes from a desk to several city blocks. We make public a ROS implementation.

本文介绍了ORB-SLAM，一种基于特征的单目SLAM系统，能够实时运行，在小型和大型、室内和室外环境中运行，并具有鲁棒性、宽基线环路闭合和重定位等特点。

ORB-SLAM：一种多功能、精准的单目SLAM系统