ORB-SLAM2: an Open-Source SLAM System for Monocular, Stereo and RGB-D Cameras (1610.06475v2)

Abstract: We present ORB-SLAM2 a complete SLAM system for monocular, stereo and RGB-D cameras, including map reuse, loop closing and relocalization capabilities. The system works in real-time on standard CPUs in a wide variety of environments from small hand-held indoors sequences, to drones flying in industrial environments and cars driving around a city. Our back-end based on bundle adjustment with monocular and stereo observations allows for accurate trajectory estimation with metric scale. Our system includes a lightweight localization mode that leverages visual odometry tracks for unmapped regions and matches to map points that allow for zero-drift localization. The evaluation on 29 popular public sequences shows that our method achieves state-of-the-art accuracy, being in most cases the most accurate SLAM solution. We publish the source code, not only for the benefit of the SLAM community, but with the aim of being an out-of-the-box SLAM solution for researchers in other fields.

• The paper introduces a versatile SLAM system that supports monocular, stereo, and RGB-D inputs to enable real-time localization and mapping.
• It leverages both local and full bundle adjustment to achieve accurate, metric-scale trajectory estimation and minimize scale drift.
• The open-source release outperforms leading benchmarks on KITTI, EuRoC, and TUM RGB-D, promoting further SLAM research and development.

ORB-SLAM2: An Open-Source SLAM System for Monocular, Stereo and RGB-D Cameras

Simultaneous Localization and Mapping (SLAM) has significantly advanced in the past two decades, particularly within the Computer Vision and Robotics communities. "ORB-SLAM2: an Open-Source SLAM System for Monocular, Stereo and RGB-D Cameras" by Raúl Mur-Artal and Juan D. Tardós, presents a highly versatile and efficient SLAM system, notable for its ability to work with monocular, stereo, and RGB-D cameras. The system introduces significant innovations in real-time SLAM for various environments and scenarios.

Key Contributions and Features

1. Comprehensive SLAM Capability:
   • ORB-SLAM2 supports monocular, stereo, and RGB-D inputs, making it adaptable to various applications and environments.
   • It integrates substantial functionalities, such as map reuse, loop closing, and relocalization—all crucial for creating accurate and reliable SLAM solutions.
2. Real-Time Operation:
   • The system operates in real-time on standard CPUs, efficiently handling sequences from handheld indoor scenarios to drone operations and vehicular environments.
3. Bundle Adjustment (BA):
   • The backend relies on BA for monocular and stereo observations, allowing for precise trajectory estimation with metric scale.
   • It incorporates both local and full BA to achieve globally consistent mapping and pose estimation.
4. Lightweight Localization Mode:
   • ORB-SLAM2 includes a mode that leverages visual odometry for unmapped regions, ensuring zero-drift localization by matching visual map points.
5. Open-Source Availability:
   • By publishing the source code, the authors aim to provide an out-of-the-box solution for the research community, promoting advancements across various fields.

Empirical Evaluation and Results

The system's performance was exhaustively evaluated on 29 popular public sequences, demonstrating state-of-the-art accuracy. The evaluation included datasets such as KITTI, EuRoC, and TUM RGB-D, comparing ORB-SLAM2 to other leading SLAM approaches.

1. KITTI Dataset:
   • ORB-SLAM2 outperformed the Stereo LSD-SLAM in terms of absolute translation RMSE and relative translation and rotation errors.
   • The system successfully detected loops in urban sequences and provided accurate localization even in challenging environments like highways.
2. EuRoC Dataset:
   • ORB-SLAM2 achieved exemplary results in high-speed MAV sequences, overcoming issues related to motion blur and tracking loss.
   • It operated with high precision, showcasing robustness in various environmental conditions.
3. TUM RGB-D Dataset:
   • ORB-SLAM2 demonstrated superior accuracy compared to ElasticFusion, Kintinuous, DVO-SLAM, and RGB-D SLAM in multiple indoor scenarios.
   • The results highlighted the efficacy of BA over ICP and other direct methods for accurate camera localization.

Technical Implications

ORB-SLAM2's development underscores critical advancements in SLAM's theoretical and practical aspects. The system's ability to integrate depth information from single frames allows for straightforward bootstrapping and minimizes scale drift—a significant improvement over monocular-only systems. Additionally, the inclusion of lightweight localization tailored for well-mapped areas holds promise for applications requiring long-term stability and precision, such as augmented reality and virtual reality.

Future Developments

Potential future directions for ORB-SLAM2 include:

• Supporting Advanced Camera Systems:
  • Non-overlapping multi-camera and omnidirectional cameras could expand ORB-SLAM2's applicability in diverse and complex environments.
• Enhanced Robustness:
  • Integrating IMU data to bolster performance under severe motion blur and optimization for environments with limited texture could further extend its robustness.
• Cooperative Mapping:
  • Implementing cooperative SLAM approaches to enable multi-robot fusion and collaborative mapping.

Overall, ORB-SLAM2 stands as a significant contribution to the SLAM community, providing a flexible, accurate, and real-time capable SLAM system adaptable to various visual sensor inputs. The open-source nature of the project promotes ongoing improvements and supports the research community in further advancing the state of SLAM technology.