This paper presents ORB-SLAM2, an open-source SLAM system for monocular, stereo, and RGB-D cameras. The system supports real-time operation on standard CPUs and includes features such as map reuse, loop closing, and relocalization. It uses bundle adjustment with monocular and stereo observations to achieve accurate trajectory estimation with metric scale. The system also includes a lightweight localization mode that enables zero-drift localization by leveraging visual odometry tracks and matching to map points. Evaluation on 29 popular public sequences shows that ORB-SLAM2 achieves state-of-the-art accuracy, being the most accurate SLAM solution in most cases. The source code is published to benefit the SLAM community and to provide an out-of-the-box SLAM solution for researchers in other fields. ORB-SLAM2 is built on the monocular ORB-SLAM system and introduces several contributions, including the first open-source SLAM system for monocular, stereo, and RGB-D cameras, RGB-D results showing higher accuracy than state-of-the-art methods based on ICP or photometric and depth error minimization, and stereo results more accurate than state-of-the-art direct stereo SLAM. The system processes stereo and RGB-D inputs to estimate camera trajectory and build a map of the environment, capable of closing loops, relocalizing, and reusing its map in real-time on standard CPUs with high accuracy and robustness. The system uses ORB features for tracking, mapping, and place recognition, which are robust to rotation and scale and provide good invariance to camera auto-gain and auto-exposure, and illumination changes. It exploits stereo/depth information to synthesize stereo coordinates for extracted features, allowing the system to be agnostic of the input being stereo or RGB-D. The back-end is based on bundle adjustment and builds a globally consistent sparse reconstruction, making the system lightweight and suitable for standard CPUs. The system includes a Place Recognition module based on DBoW2 for relocalization, a covisibility graph for linking keyframes, and a lightweight localization mode for long-term and globally consistent localization. It also includes a loop closing process that detects and validates loops, and performs pose-graph optimization to correct accumulated drift. The system is evaluated on three popular datasets, including the KITTI, EuRoC, and TUM RGB-D datasets, showing superior accuracy compared to other state-of-the-art SLAM systems. ORB-SLAM2 achieves the highest accuracy in most cases, with zero-drift localization in already mapped areas and better performance than direct methods or ICP in RGB-D results. The source code is released for use by other researchers, and ORB-SLAM2 is the first open-source visual SLAM system that can work with monocular, stereo, and RGB-D inputs.This paper presents ORB-SLAM2, an open-source SLAM system for monocular, stereo, and RGB-D cameras. The system supports real-time operation on standard CPUs and includes features such as map reuse, loop closing, and relocalization. It uses bundle adjustment with monocular and stereo observations to achieve accurate trajectory estimation with metric scale. The system also includes a lightweight localization mode that enables zero-drift localization by leveraging visual odometry tracks and matching to map points. Evaluation on 29 popular public sequences shows that ORB-SLAM2 achieves state-of-the-art accuracy, being the most accurate SLAM solution in most cases. The source code is published to benefit the SLAM community and to provide an out-of-the-box SLAM solution for researchers in other fields. ORB-SLAM2 is built on the monocular ORB-SLAM system and introduces several contributions, including the first open-source SLAM system for monocular, stereo, and RGB-D cameras, RGB-D results showing higher accuracy than state-of-the-art methods based on ICP or photometric and depth error minimization, and stereo results more accurate than state-of-the-art direct stereo SLAM. The system processes stereo and RGB-D inputs to estimate camera trajectory and build a map of the environment, capable of closing loops, relocalizing, and reusing its map in real-time on standard CPUs with high accuracy and robustness. The system uses ORB features for tracking, mapping, and place recognition, which are robust to rotation and scale and provide good invariance to camera auto-gain and auto-exposure, and illumination changes. It exploits stereo/depth information to synthesize stereo coordinates for extracted features, allowing the system to be agnostic of the input being stereo or RGB-D. The back-end is based on bundle adjustment and builds a globally consistent sparse reconstruction, making the system lightweight and suitable for standard CPUs. The system includes a Place Recognition module based on DBoW2 for relocalization, a covisibility graph for linking keyframes, and a lightweight localization mode for long-term and globally consistent localization. It also includes a loop closing process that detects and validates loops, and performs pose-graph optimization to correct accumulated drift. The system is evaluated on three popular datasets, including the KITTI, EuRoC, and TUM RGB-D datasets, showing superior accuracy compared to other state-of-the-art SLAM systems. ORB-SLAM2 achieves the highest accuracy in most cases, with zero-drift localization in already mapped areas and better performance than direct methods or ICP in RGB-D results. The source code is released for use by other researchers, and ORB-SLAM2 is the first open-source visual SLAM system that can work with monocular, stereo, and RGB-D inputs.