3D Point Cloud Registration Using Multiple Kinects

Abstract

Before any reglstrJtlon can be done, I need some data to work on. Accessing the information from the Klnect Is the first step. using the libfreenect driver I acqu1re an RGB matrix and a depth matrix both 640x480 In dimension. The next step in the process of reglstratton requires having correspondences. I begin finding the correspondences by locatmg SIFT features in each of the RGB images. A quick compJrison between finger prints of all of the SIFT features gtves me a set of potential correspondences. To narrow down the large collection of correspondences I use a filtering process which I'll elaborate on below. Once the corre~pondences have been located I perform an implementation of Procrustes Analysis to align the correspondences and ultimately the point clouds.