Approximate k-d tree search for efficient ICP

A method is presented that uses an Approximate Nearest Neighbor method for determining correspondences within the Iterative Closest Point Algorithm. The method is based upon the k-d tree. The standard k-d tree uses a tentative backtracking search to identify nearest neighbors. In contrast, the Approximate k-d tree (Ak-d tree) applies a depth-first nontentative search to the k-d tree structure. This search improves runtime efficiency, with the tradeoff of reducing the accuracy of the determined correspondences. This approximate search is applied to early iterations of the Iterative Closest Point Algorithm, transitioning to the standard k-d tree for the final iterations after the change in the mean square error of the correspondences becomes sufficiently small. The method benefits both from the improved time performance of the approximate search in early iterations as well as the full accuracy of the complete search in later iterations. Experimental results indicate that the time efficiency of Ak-d tree is superior to the k-d tree and Elias for moderately large point sets. The change in the shape of the minimum potential well space is subtle, and the convergence properties are often identical. In those cases where the global minimum was not achieved, the difference in final mse was very small. In one trial, Ak-d tree converged faster to a better minimum with a smaller mse, which indicates that the use of approximate methods may be beneficial in the presence of outliers.