A Modified ICP Algorithm Based on FAST and Optical Flow for 3D Registration

This paper presents a modified Iterative Closest Point (ICP) algorithm based on a suitable selection of initial points and local optical flow to speed up registration of static scenes with high accuracy. The biggest disadvantages of using standard ICP algorithm are appropriate initialization and effective matching point step in each iteration. In the proposed modification we deal with these problems and optimize this method for Augmented Reality application. As this application uses RGB-D images sequence the changes between consecutive key-frames are small. Therefore only small subset of the source image key-points is selected using scale-space pyramid and FAST approaches. It leads to the significant reduction of the number of the processed image points. Since the point matching technique using local optical flow is applied, in each optimization step of ICP the costly point matching procedure can be abandoned. The proposed approach has been validated by the numerical examples.