Skip to main content

​
What is SLAM?

Simultaneous Localization and Mapping (SLAM) is a computational problem where a robot or autonomous system must build a map of an unknown environment while simultaneously keeping track of its location within that environment. ORB-SLAM3 is a real-time SLAM library that performs Visual, Visual-Inertial, and Multi-Map SLAM with monocular, stereo, and RGB-D cameras.

​
Core Architecture

ORB-SLAM3 consists of four main parallel threads that work together to achieve robust tracking and mapping:

Tracking

Processes every frame to compute the camera pose by matching features with the local map. Decides when to insert keyframes.

Local Mapping

Manages the local map and performs local bundle adjustment to refine the camera poses and 3D point positions.

Loop Closing

Detects loops with every new keyframe and performs pose graph optimization when a loop is found.

Map Management

Handles multiple maps, map merging, and atlas management for seamless multi-session SLAM.

​
System Components

​
ORB Feature Extraction

ORB-SLAM3 uses ORB (Oriented FAST and Rotated BRIEF) features for visual tracking:
  • Scale invariance: Multi-scale pyramid for robust feature detection
  • Rotation invariance: Oriented features handle camera rotation
  • Computational efficiency: Real-time performance on standard CPUs

​
Place Recognition

The system uses DBoW2 (Bag of Binary Words) for:
  • Loop closure detection
  • Relocalization after tracking loss
  • Map matching for multi-map scenarios

​
Graph Optimization

ORB-SLAM3 leverages g2o for non-linear optimization:
  • Local bundle adjustment in mapping thread
  • Pose graph optimization after loop closure
  • Full bundle adjustment for map refinement

​
Thread Architecture

The Tracking thread runs in the main execution thread, while Local Mapping, Loop Closing, and Viewer run in separate threads for parallel processing.

​
Key Features

Robust monocular, stereo, and RGB-D SLAM without inertial sensors. Achieves accurate 6-DOF camera pose estimation using visual features alone.
Integrates IMU measurements at frame rate for improved robustness, especially during fast motion, occlusions, or texture-poor environments.
Creates and manages multiple maps in the same session. Automatically merges maps when common areas are detected.
Supports loading previously created maps and continuing SLAM in known environments.

​
Tracking States

The system can be in different tracking states managed through the System class:
  • OK: Successfully tracking the current map
  • LOST: Tracking lost, attempting relocalization
  • RECENTLY_LOST: Tracking recently lost, using motion model
  • NOT_INITIALIZED: System not yet initialized

​
Operating Modes

​
SLAM Mode (Default)

System SLAM(strVocFile, strSettingsFile, System::STEREO, true);
// Full SLAM: tracking + mapping + loop closing
The system simultaneously builds the map and localizes the camera.

​
Localization-Only Mode

SLAM.ActivateLocalizationMode();
// Only tracking, no mapping
In localization mode, the local mapping thread is stopped. The system only tracks the camera pose without updating the map.

​
System Initialization

The System class is the main entry point defined in include/System.h:83-265: 
System(const string &strVocFile, 
       const string &strSettingsFile, 
       const eSensor sensor, 
       const bool bUseViewer = true, 
       const int initFr = 0, 
       const string &strSequence = std::string());
Parameters:
  • strVocFile: Path to ORB vocabulary for place recognition
  • strSettingsFile: YAML configuration file with camera and sensor parameters
  • sensor: Sensor mode (see Sensor Modes)
  • bUseViewer: Enable/disable visualization

​
Performance Characteristics

Accuracy

Significantly more accurate than other open-source SLAM systems across all sensor configurations.

Robustness

As robust as the best systems available in the literature, handles challenging scenarios.

Real-time

Runs in real-time on modern CPUs (e.g., Intel i7) for live applications.

Sensor Modes

Learn about different sensor configurations supported by ORB-SLAM3

Multi-Map System

Understand the Atlas and multi-map capabilities

Camera Models

Explore pinhole and fisheye camera models

IMU Integration

Deep dive into visual-inertial fusion

​
Next Steps

1

Choose Your Sensor Mode

Review the Sensor Modes page to understand which configuration matches your hardware.
2

Configure Your Camera

Set up your camera calibration file following the Camera Models guide.
3

Build the System

Follow the Building from Source guide to compile ORB-SLAM3.

Build docs developers (and LLMs) love

Get started for freeTalk to us