Dyna VIO: Real-Time Visual-Inertial Odometry with Instance Segmentation in Dynamic Environments

Current state-of-the-art visual-inertial odometry algorithms work on the assumption that the observed features are absolutely static, which leads to their weakness in dynamic environments. In this paper, we propose a novel visual-inertial odometry framework that is capable of identifying and rejecting dynamic features, called DynaVIO. First, instance information of input images are collected to detect potential dynamic objects. Then, consecutive masks of the same object are associated to maintain the information consistency of tracked features. Based on this, a dynamic probability propagation model is built to infer the state inconsistency between masks and static background, achieving fast and accurate identifications of dynamic objects. Finally, features on highly dynamic objects are rejected to avoid their influences on pose estimation tasks. Validation experiments on VIODE dataset proves that our algorithm perform better than other state-of-the-art methods.