Consistent and Optimal Solution to Camera Motion Estimation.

Area of Science:

• Computer Vision
• Robotics
• Photogrammetry

Background:

• Camera motion estimation from 2D point correspondences is crucial for computer vision tasks.
• Existing methods often rely on the epipolar constraint, which may not be optimal in the maximum likelihood sense.

Purpose of the Study:

• To develop a novel, statistically optimal algorithm for camera motion estimation.
• To address limitations of existing methods by directly modeling the measurement error.

Main Methods:

• Formulated a maximum likelihood (ML) problem directly from the measurement model.
• Proposed a two-step algorithm: bias elimination for noise variance estimation and Gauss-Newton iteration on a manifold for refinement.
• Proved consistency and asymptotic efficiency of the proposed estimator.

Main Results:

• The proposed estimator achieves consistency and asymptotic efficiency, matching the Cramer-Rao lower bound.
• Demonstrated linear time complexity, advantageous for dense point correspondences.
• Experimental results show superior accuracy and speed compared to state-of-the-art methods on synthetic and real data.

Conclusions:

• The novel two-step algorithm provides a statistically optimal and computationally efficient solution for camera motion estimation.
• This method offers significant advantages for applications with dense point correspondences.
• The findings advance the state-of-the-art in camera motion estimation accuracy and performance.