Real-Time 3D Mapping with Visual-Inertial Odometry Pose Coupled with Localization in an Occupancy Map

Abstract

Recent research has shown that visual and inertial measurements can serve as a powerful, robust and accurate odometry source when processed by state-of-the-art algorithms. One of the main benefits of such approach is short latency, even for on-board computers working on Miniature Autonomous Vehicles (MAV). However, depending on environmental conditions or sensor motion patterns, this type of odometry may be prone to drift or even divergence. In the presented work, it is shown that employing occupancy maps can limit such undesirable behaviour while still providing pose estimate at high frequencies. This is of particular importance for highly dynamical MAV control with limited on-board numerical capabilities.