A Fast and Robust Extrinsic Calibration for RGB-D Camera Networks

Area of Science:

• Computer Vision
• Robotics
• 3D Graphics

Background:

• RGB-Depth (RGB-D) camera networks are crucial for applications like object tracking and 3D reconstruction.
• Calibrating sparse camera networks is challenging due to limited shared features across views.
• Cost-efficiency and visibility in practical applications necessitate the use of sparsely placed cameras.

Purpose of the Study:

• To develop a novel algorithm for accurate and rapid calibration of arbitrary numbers of RGB-D cameras in a network.
• To address the challenges of calibrating sparse camera networks with wide camera separations.
• To improve the robustness and generalization of camera extrinsic calibration.

Main Methods:

• Utilized a spherical calibration object to establish view correspondences, outperforming planar objects.
• Systematically evaluated rigid, polynomial, and manifold regression transformations for view mapping, moving beyond rigid transformations.
• Reformulated bundle adjustment to minimize global 3D reprojection error for fine-tuning estimates.
• Implemented a scalable client-server architecture for computational efficiency.

Main Results:

• The proposed algorithm accurately and rapidly calibrates geometric relationships in RGB-D camera networks.
• Spherical calibration objects proved superior for establishing correspondences in widely separated cameras.
• Manifold regression demonstrated robust generalization for view transformation.
• A five-camera system was calibrated in minutes using commodity PCs.

Conclusions:

• The novel algorithm provides an accurate, efficient, and robust solution for calibrating sparse RGB-D camera networks.
• The framework's scalability and speed make it suitable for practical vision and graphics applications.
• The use of spherical objects and advanced transformation models enhances calibration accuracy and generalization.