Automatic Camera Calibration Using Multiple Sets of Pairwise Correspondences.

Area of Science:

• Computer Vision
• Robotics
• Photogrammetry

Background:

• Establishing camera networks often requires extensive view overlap for calibration.
• Traditional methods struggle with limited overlap, necessitating triple correspondences.
• Uncalibrated node addition is challenging with sparse data.

Purpose of the Study:

• To develop a robust method for integrating uncalibrated cameras into existing networks.
• To overcome the limitations of triple correspondence methods in challenging scenarios.
• To improve calibration and motion estimation with pairwise correspondences.

Main Methods:

• An 11-point minimal solution for intrinsic and extrinsic camera calibration using pairwise correspondences.
• A novel inlier selection framework extending RANSAC for multi-dataset sampling.
• Validation across scenarios with limited view overlap.

Main Results:

• Successfully added uncalibrated nodes to camera networks using only pairwise correspondences.
• Achieved accurate calibration and motion estimation for moving cameras within networks.
• Enhanced Structure-from-Motion models with views exhibiting limited overlap.

Conclusions:

• The proposed method offers a significant advantage over triple correspondence techniques when overlap is limited.
• This approach broadens the applicability of camera network calibration and Structure-from-Motion.
• Enables robust integration of new nodes and views in challenging environments.