Unsupervised Deep Visual-Inertial Odometry with Online Error Correction for RGB-D Imagery

Area of Science:

• Robotics
• Computer Vision
• Machine Learning

Background:

• Real-world robotic systems face novel sensory inputs challenging existing vision-aided localization methods.
• Deep learning approaches for vision-aided localization often struggle with unseen data.
• Accurate localization is critical for autonomous navigation and robotic applications.

Purpose of the Study:

• To develop a robust visual-inertial odometry (VIO) system capable of handling novel sensory states.
• To introduce online error correction (OEC) modules for improving vision-aided localization accuracy.
• To enable absolute trajectory estimation by fusing RGB-D imagery with inertial measurements without prior calibration.

Main Methods:

• Developed the Visual-Inertial-Odometry Learner (VIOLearner), an unsupervised deep neural network for VIO.
• Integrated online error correction (OEC) modules trained to rectify localization mistakes.
• Fused RGB-D imagery with inertial measurement unit (IMU) data, bypassing the need for IMU intrinsic parameters or extrinsic calibration.

Main Results:

• Demonstrated the generalizability of OEC modules across diverse datasets.
• Achieved state-of-the-art (SoA) translational localization performance on the KITTI Odometry dataset and a custom MAV dataset.
• VIOLearner successfully performed VIO without requiring IMU intrinsic parameters or camera-IMU extrinsic calibration.

Conclusions:

• The proposed VIOLearner with OEC modules offers a robust solution for vision-aided localization in challenging real-world scenarios.
• The system demonstrates superior performance compared to existing VIO, VO, and VSLAM approaches.
• This work advances the field of autonomous navigation by improving the reliability of localization systems.