Real-Time SLAM Relocalisation

Abstract

Monocular SLAM has the potential to turn inexpensive cameras into powerful pose sensors for applications such as robotics and augmented reality. However, current implementations lack the robustness required to be useful outside laboratory conditions: blur, sudden motion and occlusion all cause tracking to fail and corrupt the map. Here we present a system which automatically detects and recovers from tracking failure while preserving map integrity. By extending recent advances in keypoint recognition the system can quickly resume tracking - i.e. within a single frame time of 33 ms - using any of the features previously stored in the map. Extensive tests show that the system can reliably generate maps for long sequences even in the presence of frequent tracking failure.