Research And Application Of Co-clustering Algorithm Based On Spectral Partition

In recent years, both structure extraction and point clouds registration have become hot topics in the field of computer vision research and application. Point clouds registration is a key technology in 3D reconstruction. Considering the ground plane and buildings always consist of several planes, we propose a new method that registering large-scale point clouds together through matched planes.After learning many classics structure extraction algorhthms, this paper presents a novel co-clustering method giving consideration to both the similarity among data points and the similarity among structure model hypotheses, and then registers building point clouds. This paper has two innovations:(1) introducing the co-clustering method. Differing from the classical clustering algorithm based on concept space, such as J-Linkage and AKSWH which cluster data points or data hypotheses seperately. This paper adopts co-clustering method based on spectral partitioning to further improve the traditional structure extraction algorithms. Finally, the clustering results of both synthetic and real data show that our algorithm not only can fitting muptiple model instances but also perform more stable and accurate when precentage of noise is high. (2) Obtaining planes from point clouds accurately using co-clustering. After knowing the planes, we use the random sample method to match the planes. The transformation matrix between the point clouds is calculated from the matched planes. Experimental results not only show that the co-clustering method can extract the planar structure more accurate and robust, but also demonstrate the effectiveness of the novel registration method for building point clouds.