Three Dimensional Shape Characteristic Analysis Based On Main Manifold

As an emerging class of biometrics, human ear has drawn significant attention in recentyears. With the development of the laser scanning technology and digital geometry processingtheory, the effective description of3D point cloud has become an important research topic inthe field of ear recognition and3D model retrieval.Focusing on the ear identification and3D model retrieval, we study the ear matchingtechniques on3D point cloud and feature extraction method of point cloud unevenlydistributed in space. The major contributions of the paper are:1. We introduce a3D ear recognition system based on the local saliency and quadricprincipal manifold. First, we propose a novel method for computing saliency value ofeach point on3D ear point clouds, which is based on the Gaussian-weighted average ofthe mean curvature and can be used to sort the keypoints accordingly. Then we proposethe optimal selection of the salient key points using the Poisson Disk Sampling. Finally,we fit a surface to the neighborhood of each salient keypoint using the quadratic principalmanifold method, establishing the local feature descriptor of each salient keypoint. Theexperimental results on ear shape matching show that, compared with other similarmethods, the proposed system has higher approximation precision on shape featuredetection and higher matching accuracy on the ear recognition.2. We present a framework for the feature detection of3D point cloud. We construct2Dprincipal manifold for the3D point cloud in form of watertight mesh with sphericalhomeomorphism. By this means, the shape description of3D point cloud has beenconverted into the description of2D principal manifold evenly spread in sphericalparameter field. Then we apply translation, rotation and scale (rough alignment) to thequadratic optimized mesh to align the mesh polar axis. Finally, the ICNP (IterativeClosest Normal Point) algorithm is used to iteratively refine the transformation to bringthe two meshes into the best alignment in the sense of the least mean square error. Thealignment error is recorded as difference distance between two3D point clouds. Theexperimental results show that the proposed3D point cloud shape description based onquadratic optimized approximation of2D principal manifold is robust to noise andresolution, and can be used as the shape description for3D retrieval.