Semantic Monocular SLAM for Highly Dynamic Environments

Nikolas Brasch, Aljaz Bozic, Joe Lallemand, Federico Tombari

Abstract: Recent advances in monocular SLAM have en-abled real-time capable systems which run robustly under theassumption of a static environment, but fail in presence ofdynamic scene changes and motion, since they lack an explicitdynamic outlier handling. We propose a semantic monocularSLAM framework designed to deal with highly dynamic en-vironments, combining feature-based and direct approaches toachieve robustness under challenging conditions. The proposedapproach exploits semantic information extracted from thescene within an explicit probabilistic model, which maximizesthe probability for both tracking and mapping to rely on thosescene parts that do not present a relative motion with respect tothe camera. We show more stable pose estimation in dynamicenvironments and comparable performance to the state of theart on static sequences on the Virtual KITTI and Synthia datasets.