Deep-Learning based Approaches to Visual-Inertial Odometry for Autonomous Tracking Applications

摘要

Recent geometric approaches to visual-inertial odometry have shown impressive accuracy with real-time performance in autonomous tracking applications in several fields including virtual and augmented reality (VR & AR) as well as robotics. But these methods are still not robust to challenging conditions due to their dependence on hand-engineered features, heuristics, sensor calibration and manual synchronization (when using visual and inertial sensors). In this paper, we review the recent advances in deep learning based approaches to odometry and identify some future research directions.