Dense RGB-D SLAM with Multiple Cameras

Area of Science:

• Robotics
• Computer Vision
• 3D Reconstruction

Background:

• Multi-camera dense RGB-D SLAM offers potential for improved scene reconstruction speed and localization accuracy.
• Key challenges include sensor calibration with limited common fields of view and effective fusion of multi-sensor location data.

Purpose of the Study:

• To develop and evaluate a multi-camera dense RGB-D SLAM system, specifically a three-Kinect setup.
• To propose novel calibration methods for systems with and without inertial measurement units (IMUs).
• To enhance existing RGB-D SLAM techniques for multi-camera fusion and pose optimization.

Main Methods:

• Developed two extrinsic calibration methods: one for IMU-assisted systems using improved hand-eye calibration, and another for pure visual SLAM.
• Extended a state-of-the-art single RGB-D SLAM method to a multi-camera framework.
• Implemented independent pose tracking for each camera, using the minimal-error pose as a reference to correct others.
• Improved the deformation graph by adding a device number attribute to distinguish and process surfels from different cameras.

Main Results:

• Demonstrated accurate extrinsic calibration methods through experimental verification.
• Achieved satisfactory 3D reconstructed models using the multi-camera dense RGB-D SLAM system.
• Obtained a root-mean-square error (RMSE) of 1.55 cm for length measurements in reconstructed models, comparable to single-camera systems.

Conclusions:

• The proposed multi-camera dense RGB-D SLAM system effectively addresses calibration and sensor fusion challenges.
• The system achieves high accuracy in localization and scene reconstruction.
• The developed methods provide a robust solution for complex multi-camera SLAM applications.