Uncertainty-Driven, Point-Based Image Registration

Point-based registration is the problem of computing the transformation that best aligns two point sets, such as might be obtained using range scanners or produced by feature extraction algorithms. The Iterative Closest Points (ICP) algorithm and its variants are the most commonly used techniques for point-based registration. The ICP algorithm may be derived as the solution to a global optimization problem. A commonly-used linearization of the distance function in this optimization problem produces a useful approximation to the covariance matrix of the ICP-estimated transformation parameters. Two recent algorithms exploit this covariance matrix to improve ICP registration. One uses the covariance matrix to sample the correspondences so that the estimate is well-constrained in all directions in parameter space. A second uses the covariance matrix to guide a region-growing and model-selection technique that “grows” accurate estimates from low-order in ingles estimates that are only accurate in small image regions. Both show substantial improvements over standard ICP on challenging alignment problems.