Robotics
Loop closure
Definition
Loop closure is the recognition and estimation of a constraint created when a robot revisits a previously mapped place. Adding that long-range constraint can correct accumulated drift in a localization or mapping trajectory.
Also known as: Loop closing, Loop-closure detection
Updated
Reconnecting the trajectory
Odometry accumulates error as a robot moves. When the robot returns to a known place, a loop-closure system tries to recognise the revisited scene and estimate the relative pose between the old and current observations. That creates a constraint between states separated by many time steps.
In a factor graph, this constraint adds an edge that closes a cycle in the trajectory. Optimising the graph can distribute the previous drift across many poses instead of applying a discontinuous correction only at the current pose.
Place recognition needs geometric checking
Candidate loops can be proposed from camera appearance, lidar geometry, learned descriptors, or combinations of sensors. The FAB-MAP 2.0 paper describes probabilistic appearance-based place assignment and reports loop-closure detection over a 1,000 km dataset. Its scale and timing figures are results for that system and dataset.
Similar-looking places can create false matches, so a candidate normally needs temporal consistency, geometric verification, or both. ORB-SLAM3 reports using place recognition to merge maps after revisiting mapped areas. That is one system design; loop closure can also operate without visual features.
False closures can be worse than missed ones
A missed loop leaves drift uncorrected. A false loop can deform an otherwise consistent map by asserting that two different places are the same. Repetitive corridors, lighting changes, moving objects, seasonal changes, and viewpoint differences all make recognition harder.
Loop closure is therefore not simply detecting visual similarity. A complete evaluation should report false positives, recall, pose accuracy, and the effect on the final map. The front end that proposes a match and the back end that incorporates it into simultaneous localization and mapping are separate failure points.
Sources
Related terms
Simultaneous localization and mapping
Simultaneous localization and mapping is the joint estimation of a robot's state and a map of its environment from sensor observations. It is commonly abbreviated SLAM.
Odometry
Odometry estimates changes in a robot's position and orientation from motion measurements over time. Its accumulated pose provides a local reference that can drift as measurement errors build up.
Factor graph
A factor graph is a bipartite graph whose variable nodes represent unknown quantities and whose factor nodes represent functions involving subsets of those variables. Robotics systems use this structure to combine local motion, sensor, and prior constraints in estimation and optimisation problems.
Pose estimation
Pose estimation determines the position and orientation of an object or robot relative to a reference frame. For a rigid body in three-dimensional space, a full pose has three translational and three rotational degrees of freedom.