| As an important research field in computer vision and computer graphics,feature detection and matching are the basis of several tasks,including object detection,tracking and recognition.In contrast to traditional multimedia data,such as image,audio and video,point cloud captures geometric information of an object,and is not affected by rotation,scaling and illumination.Therefore,point cloud is able to describe real-world scenes in a more comprehensive and truthful way.Thanks to the rapid development of point cloud acquisition devices and the continuous improvement of computing capabilities in recent years,research on three-dimensional models has arisen as a focus in the computer vision and computer graphics field.However,feature detection,description and shape registration of three-dimensional non-rigid models is still a challenging problem that needs great efforts.Therefore,this paper conducts theoretical and technical research on the above tasks,and has made progress in the key point detection and description,salient region detection and shape registration,as follows:First,considering the problem that current algorithms fail to deal with the local surface deformations in key point detection and description,a novel local surface based three-dimensional non-rigid surface description method is proposed,including a key point detector named persistence-based Heat Kernel Signature(pHKS),and a local surface descriptor named Heat Propagation Strips(HeaPS).Experiments on several public data sets showed that the proposed method has a superior performance in terms of distinctiveness and descriptiveness and is highly robust.Second,in the problem of salient region detection,the Scale Space Clustering Evolution(SSCE)algorithm is proposed to extract the areas with rich geometric features on three-dimensional non-rigid shapes.The key feature of this algorithm is that it has the ability to extract all the salient regions on a shape and does not depend on any manual intervention or prior knowledge.At the same time,the algorithm remains robust under various model disturbances.Experiments on several public datasets showed that the algorithm has a superior performance than the state-of-the-art in terms of saliency and repeatability.To solve the problems of low accuracy and high complexity in the current three-dimensional non-rigid model registration,two algorithms are proposed with respect to the different representations of non-rigid transformations,including a Hierarchical Shape Matching(HSM)method and a Balanced Functional Shape Matching(BFSM)method.Both of them try to reduce the complexity of the problem,where the former reduces the searching space for point matching,and the latter generalizes point matching to functional matching,and represents non-rigid transformations as a matrix.Experiments on several public data sets showed that both of them have an excellent performance in terms of accuracy and computational complexity. |
