RGBD-Inertial Trajectory Estimation and Mapping for Ground Robots.

Area of Science:

• Robotics
• Computer Vision
• Sensor Fusion

Background:

• Simultaneous Localization and Mapping (SLAM) is crucial for autonomous robots.
• Laser-based SLAM has limitations; camera-based SLAM offers cost and robustness advantages.
• RGBD sensors combine dense visual data with depth information, offering enhanced SLAM capabilities.

Purpose of the Study:

• To develop and evaluate a novel visual SLAM system, VINS-RGBD, by fusing RGBD and IMU data.
• To extend the VINS-Mono system to leverage depth data for improved initialization and Visual Inertial Odometry (VIO).
• To integrate real-time mapping with loop closing using subsampled depth data and octree filtering.

Main Methods:

• Extension of the VINS-Mono open-source software to incorporate RGBD sensor data.
• Utilizing depth information during system initialization and the VIO phase.
• Implementation of a real-time mapping system employing depth data subsampling and octree filtering for efficient map representation and loop closure detection.

Main Results:

• The proposed VINS-RGBD system demonstrates superior performance compared to VINS-Mono.
• The VINS-RGBD approach shows significant advantages over ORB-SLAM2, which failed in the tested application scenarios.
• Extensive experiments across diverse environments with hand-held, wheeled, and tracked robots validate the system's robustness and effectiveness.

Conclusions:

• Fusing RGBD and IMU data provides a robust and effective solution for ground robot SLAM.
• VINS-RGBD offers a significant advancement over existing methods like VINS-Mono and ORB-SLAM2 for specific robotic applications.
• The developed system, along with provided datasets and software, facilitates further research and development in visual-inertial SLAM.