Enhancing semi-dense monocular vSLAM used for multi-rotor UAV navigation in indoor environment by fusing IMU data

We propose the enhancement of the modern vision-based monocular simultaneous localization and mapping (vSLAM) method, e.g. LSD-SLAM, used for compact multi-rotor UAV indoor navigation, by fusing inertial measurement unit (IMU) data with camera images. We suggest removing the cost-expensive loop-closure optimization algorithm from the vSLAM pipeline and replacing it with the computationally efficient flow estimation procedure based purely on IMU data. The input IMU flow is being processed by the Extended Kalman Filter (EKF) based techniques for localization purposes and used further in LSD-SLAM algorithm for UAV pose estimation. We evaluate the proposed algorithm using the modeled indoor environment originally used for RoboCup Rescue Simulation League 2013 competition and "hector_quadrotor" - commonly used in modelling simulated UAV model. Evidently, implementation of the suggested enhancement results in significant drop-down of the runtime and leads to obtaining maps and trajectories of higher accuracy.