PhD Thesis

Solving the Nearest Rotation Matrix Problem in Three and Four Dimensions with Applications in Robotics

Work default illustration




  • Started: 11/05/2018
  • Thesis project read:


Since the map from quaternions to rotation matrices is a 2-to-1 covering map, it cannot be smoothly inverted. As a consequence, it is sometimes erroneously assumed that all inversions should necessarily contain singularities that arise in the form of quotients where the divisor can be arbitrarily small. This misconception was clarified when we found a new division-free conversion method. This result triggered the research work presented in this thesis.

At first glance, the matrix to quaternion conversion does not seem to be a relevant problem. Actually, most researchers consider it as a well-solved problem whose revision is not likely to provide any new insight in any area of practical interest. Nevertheless, we show in this thesis how solving the nearest rotation matrix problem, in Frobenius norm, can essentially be reduced to a matrix to quaternion conversion. Many problems, such as hand-eye calibration, camera pose estimation, location recognition, image stitching, etc. require finding the nearest proper orthogonal matrix to a given matrix. Thus, the matrix to quaternion conversion becomes of paramount importance.

While a rotation in 3D can be represented using a quaternion, a rotation in 4D can be represented using a double quaternion. As a consequence, the computation of the nearest rotation matrix in 4D, using our approach, essentially follows the same steps as in the 3D case. Although the 4D case might seem of theoretical interest only, we show in this thesis its practical relevance thanks to a little known mapping between 3D displacements and 4D rotations.

In this thesis we focus our attention in obtaining closed-form solutions, in particular those that only require the four basic arithmetic operations because they can easily be implemented on microcomputers with limited computational resources. Moreover, closed-form methods are preferable for at least two reasons: they provide the most meaningful answer because they permit symbolically analyzing the influence of each variable on the result; and their computational cost, in terms of arithmetic operations, is fixed and assessable beforehand. We have actually derived closed-form methods specifically tailored for solving the hand-eye calibration and the pointcloud registration problems which outperform all previous approaches.

The work is under the scope of the following projects:

  • MdM: Unit of Excellence María de Maeztu (web)