Relocaliser

Relocaliser is a component used in robotics and computer vision to re-establish the pose of a device within a known map after tracking has been lost. Its purpose is to determine where the current observation fits in a pre-built map and to provide an initial pose estimate that allows simultaneous localization and mapping (SLAM) or navigation to resume.

In practice, relocalisation treats the problem as global localization: given the latest sensor data, often from a camera, LiDAR, or sensor fusion, the system searches for the device’s location within the map. The relocaliser may use image or point-cloud features to match the current view to a database of map elements, and then compute a pose using geometric verification such as PnP with RANSAC or pose-graph optimization. Some approaches are learning-based, predicting pose or selecting plausible pose candidates directly from observations.

Common methods include place recognition with feature descriptors (such as bag-of-words, neural descriptors, or learned embeddings) to identify candidate keyframes, followed by geometric refinement. In many SLAM implementations, a relocaliser is invoked when tracking degrades or is lost, providing a re-initialization that allows the system to recover after limitations like occlusions or rapid motion.

Applications of relocalisers span mobile robotics, autonomous driving, and augmented reality, where robust re-localization is essential for continuous operation in changing environments. Challenges include substantial appearance changes, dynamic objects, varying illumination, scale drift, and the need to balance accuracy with real-time performance.