Robotics

V-SLAM

2007ActivePublished: 21 October 2026Updated: 21 October 2026Published

Key innovation

Enabled simultaneous estimation of camera trajectory and 3D map of the environment from a video stream alone, removing the need for expensive laser scanners or external localization infrastructure.

How it works

A typical V-SLAM pipeline splits into a frontend and a backend, usually running in separate threads. The frontend (tracking) extracts features or photometric gradients from successive frames and estimates camera motion between them (visual odometry). Selected keyframes are passed to the backend (mapping), which optimizes a consistent map of 3D points and camera poses, typically via bundle adjustment minimizing reprojection error (feature-based) or photometric error (direct). Loop-closure detection recognizes previously visited places (e.g. via bag-of-words), after which pose-graph optimization corrects accumulated drift. A relocalization mechanism recovers the pose after tracking loss. Visual-inertial variants tightly couple IMU measurements, improving scale estimation and stability.

Problem solved

An autonomous robot moving through an unknown environment must simultaneously know where it is and understand the surrounding structure — two mutually dependent tasks. V-SLAM solves this chicken-and-egg problem using vision alone, removing dependence on GPS (unavailable indoors), expensive LiDAR and pre-built maps.

Components

Frontend (tracking / visual odometry)Real-time motion estimation

Processes successive frames, detecting features or photometric gradients and estimating relative camera motion between frames.

Official

Backend (mapping / optimization)Map consistency and accuracy

Optimizes a globally consistent 3D map and camera poses, typically via bundle adjustment or pose-graph optimization.

Official

Loop-closure detectionLong-term drift reduction

Recognizes previously visited places and triggers drift correction via pose-graph optimization.

Official

RelocalizationRobustness to tracking loss

Recovers the camera pose relative to the existing map after tracking loss.

Official

Implementation

Reference implementations

ORB-SLAM2

C++ · Raúl Mur-Artal / Juan D. Tardós

Official

ORB-SLAM3

C++ · UZ-SLAMLab (University of Zaragoza)

Official

LSD-SLAM

C++ · TUM Computer Vision Group

Official

DSO (Direct Sparse Odometry)

C++ · Jakob Engel

Official

Evolution

Original paper · 2007 · IEEE TPAMI 2007 · Andrew J. Davison

MonoSLAM: Real-Time Single Camera SLAM

Andrew J. Davison, Ian D. Reid, Nicholas D. Molton, Olivier Stasse

2003

MonoSLAM — first real-time monocular SLAM

Inflection point

Davison demonstrates real-time single-camera localization and mapping (ICCV 2003); journal version in IEEE TPAMI 2007.

2007

PTAM — separating tracking and mapping into parallel threads

Inflection point

Klein and Murray (ISMAR 2007) introduce keyframe-based bundle adjustment on a separate thread, enabling accurate real-time mapping.

2014

LSD-SLAM — direct, semi-dense method

Engel, Schöps and Cremers (ECCV 2014) optimize photometric error instead of features, producing a semi-dense map.

2015

ORB-SLAM — versatile feature-based SLAM

Inflection point

Mur-Artal, Montiel and Tardós (IEEE T-RO 2015) unify tracking, mapping, relocalization and loop closing on ORB features.

2016

DSO — Direct Sparse Odometry

Engel, Koltun and Cremers combine the direct approach with sparse photometric optimization (arXiv 2016, IEEE TPAMI 2018).

2017

ORB-SLAM2 — monocular, stereo and RGB-D

Mur-Artal and Tardós (IEEE T-RO 2017) extend the system to stereo and RGB-D cameras as open source.

2020

ORB-SLAM3 — visual-inertial and multi-map

Adds tight visual-inertial fusion and a multi-map (Atlas) system for improved robustness.

Parallelism

Parallelism level

Partially parallel

Key PTAM innovation: splitting tracking and mapping into separate threads running in parallel. Bundle adjustment is costly and usually asynchronous relative to real-time tracking.

Scope

Inference