This paper describes a method for continuous 3D registration of an object using a 3D sensor and model of the object,
significantly speeding up an iterative alignment method by using a 2D array cache. The cache stores local subtrees in a
kd-tree search to initialize the processing of subsequent data. The cache is spatially structured to match the projection of
the 3D space into the sensor’s field of view, and automatically adapts when points cross discontinuities or occlusion
boundaries. Experiments in a simulated 3D tracking and relative maneuvering scenario demonstrate the computational
speedup benefits of local subtree caching.