ORB-SLAM3: An Accurate Open-Source Library for Visual, Visual-Inertial and Multi-Map SLAM

Abstract: This paper presents ORB-SLAM3, the first system able to perform visual, visual-inertial and multi-map SLAM with monocular, stereo and RGB-D cameras, using pin-hole and fisheye lens models. The first main novelty is a feature-based tightly-integrated visual-inertial SLAM system that fully relies on Maximum-a-Posteriori (MAP) estimation, even during the IMU initialization phase. The result is a system that operates robustly in real-time, in small and large, indoor and outdoor environments, and is 2 to 5 times more accurate than previous approaches. The second main novelty is a multiple map system that relies on a new place recognition method with improved recall. Thanks to it, ORB-SLAM3 is able to survive to long periods of poor visual information: when it gets lost, it starts a new map that will be seamlessly merged with previous maps when revisiting mapped areas. Compared with visual odometry systems that only use information from the last few seconds, ORB-SLAM3 is the first system able to reuse in all the algorithm stages all previous information. This allows to include in bundle adjustment co-visible keyframes, that provide high parallax observations boosting accuracy, even if they are widely separated in time or if they come from a previous mapping session. Our experiments show that, in all sensor configurations, ORB-SLAM3 is as robust as the best systems available in the literature, and significantly more accurate. Notably, our stereo-inertial SLAM achieves an average accuracy of 3.6 cm on the EuRoC drone and 9 mm under quick hand-held motions in the room of TUM-VI dataset, a setting representative of AR/VR scenarios. For the benefit of the community we make public the source code.

• The paper introduces a novel SLAM system that integrates visual, visual-inertial, and multi-map approaches using MAP estimation for enhanced accuracy.
• It achieves sub-decimeter precision through efficient initialization and keyframe-based bundle adjustment, validated on EuRoC and TUM-VI datasets.
• The open-source library supports diverse camera models and environments, making it ideal for applications in AR, robotics, and large-scale mapping.

ORB-SLAM3: An Accurate Open-Source Library for Visual, Visual-Inertial and Multi-Map SLAM

Abstract

ORB-SLAM3 is a comprehensive system that unifies visual, visual-inertial, and multi-map SLAM using monocular, stereo, and RGB-D cameras while supporting pin-hole and fisheye lens models. It introduces robust visual-inertial SLAM reliant on Maximum-a-Posteriori (MAP) estimation, enhancing accuracy by two to ten times over previous methods. A novel multiple map system aids in maintaining precision across sessions, even under challenging conditions. ORB-SLAM3's improved recall in place recognition ensures seamless map merging and superior accuracy.

Introduction

ORB-SLAM3 extends the capabilities of visual SLAM by integrating data from various camera types and inertial sensors, enabling comprehensive, highly accurate mapping. Unlike traditional visual odometry (VO), ORB-SLAM3 uses keyframe-based bundle adjustment (BA) for optimal mapping, leveraging short, mid, and long-term data associations.

ORB-SLAM3 further integrates advanced visual-inertial initialization, ensuring precise mapping in environments with poor visual features. A multi-map feature allows revisiting and merging of previously mapped areas, supporting complex, large-scale environments.

System Architecture

Figure 1: Main system components of ORB-SLAM3.

Components

1. Tracking: Processes sensor inputs to estimate current frame pose, performing real-time adjustments using MAP estimation for monocular and stereo streams.
2. Local Mapping: Refines the active map by incorporating new keyframes and optimizing through local BA.
3. Loop and Map Merging: Continuously checks for loop closures and map merging opportunities using a novel place recognition system, enhancing data associations.

Key Innovations

Visual-Inertial SLAM

ORB-SLAM3 provides tight coupling of visual and inertial data by fully incorporating MAP estimation in its initialization phase. The initialization process is efficient, achieving a 5% scale error with as little as 2 seconds of data and converging to 1% within 15 seconds.

Multi-Map System

The robust multi-map system allows ORB-SLAM3 to operate seamlessly across sessions, aligning maps from different sessions using improved recall place recognition. This enables precise and efficient relocalization and loop closure.

Performance Analysis

ORB-SLAM3 significantly outperforms existing systems in the EuRoC dataset, achieving exceptional accuracy across all configurations. The robust initialization and data association strategies contribute to achieving sub-decimeter accuracy even in challenging AV/VR scenarios on TUM-VI benchmark datasets.

Figure 2: Colored squares represent the RMS ATE for ten different executions in each sequence of the EuRoC dataset.

Computational Requirements

ORB-SLAM3 achieves its robust and accurate performance with real-time capabilities. The mapping and tracking algorithms leverage the power of modern CPU architectures, ensuring operability on standard hardware configurations.

Conclusion

ORB-SLAM3 defines a new standard in SLAM accuracy and robustness, with its integrated use of visual and inertial data and innovative multi-map capabilities. It supports precise localization and mapping across varied environments, making it ideal for applications in augmented reality (AR), robotics, and beyond. ORB-SLAM3's open-source release ensures wide accessibility and fosters further advancements in SLAM technologies.