RotICP: A Spherical-Based Distance Algorithm for Rotation Search and Point Cloud Registration

The Iterative Closest Point (ICP) algorithm is one of the most widely used point cloud registration methods with a well known problem of requiring a good initial guess to reach a satisfactory solution. In particular, the ICP may encounter issues with large initial differences between point clouds, especially regarding the rotation. We contribute a new iterative algorithm, named RotICP, which solves the challenging simultaneous rotation and correspondence search registration problem in a vastly more robust way than compared methods. The RotICP handles rotations in a natural manner using a proposed spherical-based distance measure which captures an intuitive notion of 'rotational closeness'. We additionally prove a theorem that RotICP converges monotonically to a local minima and compare it with the classical ICP and one state-of-the-art registration algorithm on numerous simultaneous rotation and correspondence search registration problems including both synthetic and real-world data. The results show that RotICP not only handles large rotations significantly better than the other methods, but also converges in significantly less iterations than ICP while achieving state-of-the-art accuracy.