A Learning-based Approach to Cover Short-term Camera Failure in a Monocular Visual Inertial Odometry System

摘要

Localization and navigation have multiple established methods in sensing and fusion algorithms.Visual Inertial Odometry(VIO)has drawn attention recently,however,one disadvantage is camera susceptibility to disturbances such as fast motions,and moving objects.Existing researchers usually test their algorithms assuming good camera performance.In this research,we propose a learning-based method to estimate pose during brief periods of camera failure or occlusion.A Long Short-Term Memory(LSTM)network is trained during periods of good camera operation and,once trained,the LSTM provides an alternative pose estimate,available as soon as camera failure is detected.We tested our algorithm by removing the visual inputs and comparing Kalman Filter,IMU-only,and pre-trained LSTMs results.The results indicate the implemented LSTM increased the positioning accuracy by 76.2% and orientation accuracy by 26.5%.