Research On Similarity Measurement Of 3D Models

The similarity measurement and matching of 3D models is the key point of computer vision, artificial intelligence, image processing and etc. It is widely used in CAD, integrated circuit design, robots design, digital city, medical diagnoses and etc. To find out an effective and efficient 3D descriptor for 3D similarity measurement, and further utilize it in 3D model matching, searching and recognizing, is the main goal of this thesis.The thesis first introduces the common pre-processing methods through the whole procedure of 3D matching. They include discretization, spherical parameterization and normalization (movement, scale, transformation). These pre-processes are mainly used for extracting the 3D descriptors based on geometrical information. Among the latest algorithms in this area, the algorithm based on Spherical Harmonic Transform is one of the best algorithms because of its rotation-invariance, high efficiency, and good discrimination.However, the algorithm based on Spherical Harmonic Transform has information loss. To solve this problem, an ameliorate algorithm is presented in this thesis. The algorithm adds a coordinate to record distance from points on the surface of the model to the center of gravity. Therefore it's able to distinguish 3D models with partial rotations which original Spherical Harmonic method can not. Then, a novel algorithm based on force field model is proposed. 3D models as surface particle sets are assumed, and the force interactions between particles are computed to constitute a spherical descriptor. During the force computation, the mass of the particle is defined to maintain local information which improves the discrimination of the spherical descriptor. The introduction of force field model also solves the stubborn problems constantly occurring under general spatial geometric view. The experimental results show that both the algorithms are valid for 3D shape matching with high efficiency.The thesis then researched 3D matching algorithms based on topological information. Skeleton and Reeb graph are introduced. Afterward, the thesis works on segmentation algorithms based on topological information and brings on an algorithm based on Reeb graph and watershed method. The algorithm imports curvature function to reserve local information, and simplifies the segmentation through merging of the meshes. The algorithm also further constructs a connected topological tree, which can be used for topological matching.At last, the thesis tries an algorithm to combine the geometrical information and topological information. The key point of this algorithm is segmentation of 3D models to visual parts. In the thesis, we take human models as examples, and utilize the algorithm in recognizing human models and data mining. The experimental results show the validity of the algorithm in recognizing joint models (with joints rotatable).