Relocalization

Figuring out where a robot is and which way it’s facing on an existing map, after tracking is lost or it restarts.

Relocalization is a step within SLAM (Simultaneous Localization and Mapping) and visual localization: when tracking is lost due to fast motion, occlusion, or a lighting change, or when a robot restarts or is picked up and moved elsewhere (the so-called “kidnapped robot” problem), the system must recover its pose on an already-built map using only its current observation. A classic approach, used in ORB-SLAM, retrieves the keyframe most similar to the current image with a bag-of-words model (DBoW2), performs feature matching, and then solves for the camera pose using PnP combined with RANSAC. Among learned approaches, PoseNet, from Cambridge in 2015, was the first to regress a 6-degree-of-freedom camera pose directly from a single image using a convolutional network. Relocalization uses techniques similar to loop closure detection, but with a different purpose: loop closure corrects accumulated drift, while relocalization recovers a lost position.

ExampleA robot vacuum is picked up and set down in a different room; it compares the current view against its stored map, relocalizes, and resumes cleaning.

Also called

Camera Relocalization, Re-Localization

Related

Simultaneous Localization and Mapping · Loop Closure Detection · Visual Place Recognition · Perspective-n-Point · Random Sample Consensus · Adaptive Monte Carlo Localization

Sources

ORB-SLAM: a Versatile and Accurate Monocular SLAM System (arXiv 1502.00956)
PoseNet: A Convolutional Network for Real-Time 6-DOF Camera Relocalization (arXiv 1505.07427)
UZ-SLAMLab/ORB_SLAM3 (GitHub)