Practical fusion algorithms for rotation matrices: A comparative study

Many computer vision, sensor fusion, and robotic applications require the estimation of a 3 × 3 rotation matrix from a set of measured or computed 3 × 3 noisy rotation matrices. This article classifies solution methods into three categories: nonlinear least squares, linear optimal, and linear suboptimal algorithms. Their performance is compared through simulation studies. It is shown that the linear suboptimal algorithms proposed in this article have an accuracy comparable to that of the optimal algorithms and are about five times faster. Furthermore, a particular nonlinear optimization algorithm is presented that has computational complexity similar to that of the linear optimal procedures. © 1992 John Wiley & Sons, Inc.