Robust Tightly-Coupled Visual-Inertial Odometry with Pre-built Maps in High Latency Situations

Area of Science:

• Robotics and Computer Vision
• Simultaneous Localization and Mapping (SLAM)

Background:

• Real-time mobile localization is challenging due to latency and low-frequency global localization services.
• Existing visual-inertial odometry (VIO) systems struggle with drift and robustness in high-latency environments.

Purpose of the Study:

• To develop a novel monocular visual-inertial odometry system resilient to high latency.
• To leverage pre-built maps for enhanced localization accuracy and robustness on mobile devices.

Main Methods:

• Tightly coupling VIO with geometric priors from pre-built maps.
• Utilizing sparse point clouds derived from dense meshes via ray casting for feature tracking and state updates.
• Employing VIO's local accuracy to effectively remove outliers during feature tracking.

Main Results:

• The system demonstrates robust real-time performance on mobile devices, even under high latency.
• Experiments on EurocMav and simulation datasets show superior precision and robustness compared to state-of-the-art methods.
• A real-time augmented reality (AR) demo confirms the system's effectiveness with remote server-based visual localization.

Conclusions:

• The proposed VIO system effectively mitigates drift and enhances robustness by integrating pre-built map geometric priors.
• This approach is highly suitable for practical applications requiring reliable localization on remote servers.
• The system offers a significant advancement for real-time AR experiences on mobile platforms.