Estimating relative camera motion from the antipodal-epipolar constraint

Area of Science:

• Computer Vision
• Robotics
• Geometric Deep Learning

Background:

• Estimating relative camera motion is crucial for applications like autonomous navigation and 3D reconstruction.
• Existing methods often struggle with large Field-of-View (FOV) cameras or exhibit limited robustness.
• Differential techniques using antipodal points have limitations in handling larger motion ranges.

Purpose of the Study:

• To introduce a novel antipodal-epipolar constraint for decoupling camera translation and rotation.
• To develop robust algorithms for estimating camera motion using this new constraint.
• To present a novel structure-from-motion algorithm leveraging the decoupled motion estimation.

Main Methods:

• Formulation of a new antipodal-epipolar constraint for discrete camera motions.
• Development of two algorithms: one RANSAC-based and one Hough-like voting-based.
• Implementation of a structure-from-motion algorithm that bypasses explicit rotation estimation.

Main Results:

• The proposed constraint effectively decouples translational and rotational camera motion.
• Algorithms demonstrate robustness and accuracy over a wider range of motions compared to prior methods.
• The novel structure-from-motion algorithm shows competitive performance without explicit rotation estimation.

Conclusions:

• The antipodal-epipolar constraint offers a significant advancement in camera motion estimation.
• The developed algorithms provide robust and accurate solutions for real-world applications.
• This work advances the field of computer vision with improved motion estimation techniques.