Lightweight omnidirectional visual-inertial odometry for MAVs based on improved keyframe tracking and marginalization

Abstract

Due to the limited onboard resources on Micro Aerial Vehicles (MAVs), the poor real-time performance has always been an urgent problem to be solved in the practical applications of visual inertial odometry (VIO). Therefore, a lightweight omnidirectional visual-inertial odometry (LOVIO) for MAVs based on improved keyframe tracking and marginalization was proposed. In the front-end processing of LOVIO, wide field-of-view (FOV) images are captured by an omnidirectional camera, frames are tracked by semi-direct method combining of direct method with rapidity and feature-based method with accuracy. In the back-end optimization, the Hessian matrix corresponding to the error optimization equation is stepwise marginalized, so the high-dimensional matrix is decomposed and the operating efficiency is improved. Experimental results on the dataset TUM-VI show that LOVIO can significantly reduce running time consumption without loss of precision and robustness, that means LOVIO has better real-time and practicability for MAVs.