Efficient 3D registration using two-stage scheme and spatial filter

This paper proposes a fast and robust three-dimensional human face registration strategy dedicated for image-guided medical applications. In order to reduce data amount effectively while preserving data points that adequately describe curvatures, a spatial filter and a truncation procedure are introduced. The proposed registration scheme is composed of a coarse registration stage and a fine-tuning stage. In the first stage, the data amount of the image is reduced by the filter and a homogeneous selection procedure, and evolutionary computation is implemented to find global optimal rigid body transformation that aligns two 3D images. In the second stage, complete data sets before the selection procedure are exploited using the iterative closest point plus k-dimensional tree scheme. To illustrate the validity of the proposed approach, two representative problems are demonstrated. Computational results of an actual registration problem show that our approach accelerates the registration process from 356.2 seconds to 19.3 seconds when compared with a genetic algorithm and iterative closest point scheme, while the average misalignment distance reduces from 1.9508 mm to 1.1024 mm.