ORBSLAM2ORBSLAM3

ORB-SLAM2 and ORB-SLAM3 are open-source simultaneous localization and mapping (SLAM) systems developed by Raúl Mur-Artal and collaborators. They are designed for real-time camera-based mapping and localization, using ORB features and a modular, keyframe-based architecture. Both projects have been influential in robotics and computer vision, offering robust performance with relatively modest computational requirements and a focus on accuracy through multi-stage optimization, loop closure, and relocalization.

ORB-SLAM2, released in 2016, supports monocular, stereo, and RGB-D cameras. It builds a sparse map of 3D map points and keyframes, employing ORB features for tracking, local mapping, and loop closure. The system uses a bag-of-words place recognition approach for relocalization and loop detection, and performs local and global bundle adjustment to refine camera poses and map points. It runs in real time on standard CPUs, with metric scale recovered automatically for stereo and RGB-D setups, while monocular operation suffers from scale ambiguity unless initialized appropriately.

ORB-SLAM3, published later as an extension, generalizes the framework to include inertial measurement unit (IMU) data, enabling visual-inertial SLAM. It supports monocular, stereo, and RGB-D configurations with tight integration of IMU data through pre-integration and calibrated camera-IMU extrinsics. This enhances scale-consistency for monocular setups, improves robustness during aggressive motion, and strengthens relocalization and map reuse across sessions. ORB-SLAM3 retains the core concepts of ORB-SLAM2—ORB feature tracking, keyframes, place recognition, loop closure, and bundle adjustment—while expanding the sensor fusion capabilities and accuracy.

Together, ORB-SLAM2 and ORB-SLAM3 form a lineage of widely used, open-source SLAM systems, with the latter adding comprehensive visual-inertial capabilities to the proven monocular/stereo/RGB-D framework of the former.