Robust Approaches for Localization on Multi-camera Systems in Dynamic Environments

Localization of humanoid robots in real-life scenarios has to robustly tackle dynamic environments and provide coherent data and tight integration for follow-up tasks. However state-of-the-art solutions, like ORBSlam2 [1], lack this ability. In this work we present two adaptations of ORBSlam2 for a multi-camera setup on the DLR Rollin' Justin System, one distributed multi-slam and one combined single-process system. Further, we introduce the usage of pre-recorded maps with ORBSlam2 and the alignment with semantic maps for planning. We compare performance of the adaptations against and the original approach in realistic experiments and discuss advantages and disadvantages of all methods.