Adaptive Monocular Visual-Inertial SLAM for Real-Time Augmented Reality Applications in Mobile Devices.

Area of Science:

• Computer Vision
• Robotics
• Augmented Reality

Background:

• Simultaneous Localization and Mapping (SLAM) is crucial for robots, autonomous navigation, and AR.
• Fast camera pose estimation and true scale are vital for AR applications.
• Existing SLAM methods may face challenges in real-time mobile AR scenarios.

Purpose of the Study:

• To develop an adaptive monocular visual-inertial SLAM method for real-time mobile AR.
• To enhance camera pose estimation speed and accuracy.
• To dynamically adapt between visual-inertial odometry and optical-flow-based methods.

Main Methods:

• Implemented a SLAM system using visual-inertial odometry (VIO) combining camera and IMU data.
• Developed a fast optical-flow-based visual odometry (VO) for real-time pose estimation.
• Introduced an adaptive execution module to dynamically switch between VIO and optical-flow VO.

Main Results:

• Achieved an average translation root-mean-square error of 0.0617 m on the EuRoC dataset.
• Reduced average tracking time by 7.8% to 18.8% with adaptive policies.
• Demonstrated performance improvements on real mobile device sensors.

Conclusions:

• The proposed adaptive monocular visual-inertial SLAM method is effective for real-time mobile AR.
• Dynamic selection of odometry methods enhances performance and efficiency.
• The approach offers a promising solution for accurate and fast pose estimation in mobile AR.