Linearly Solving Robust Rotation Estimation.

Area of Science:

• Computer Vision
• Robotics
• Optimization

Background:

• Rotation estimation is vital for computer vision and robotics, especially in safety-critical applications.
• Traditional rotation estimation involves complex non-linear, non-convex optimization.
• Existing methods can be sensitive to noise and outliers.

Purpose of the Study:

• To reformulate rotation estimation as a linear model fitting problem.
• To develop a robust and efficient rotation estimation method.
• To demonstrate the effectiveness of the proposed method on large-scale and corrupted datasets.

Main Methods:

• Reformulation of rotation estimation as a linear model fitting problem.
• Exploration of the dual structure of rotation motion, represented as a great circle on a quaternion sphere.
• Development of a voting-based method for rotation estimation, leveraging graphics processing units (GPUs).

Main Results:

• The proposed method achieves exceptional robustness to noise and outliers.
• It enables parallel computation on GPUs for efficient processing.
• Satisfactory solutions are obtained for large-scale (10^6) and severely corrupted (99% outlier ratio) problems in under 0.5 seconds.

Conclusions:

• The novel linear framework provides a new perspective on rotation estimation.
• The voting-based GPU-accelerated method offers superior performance and robustness.
• Experimental validation confirms the effectiveness and efficiency of the proposed approach for challenging rotation estimation tasks.