Feature extraction and matching have become a popular research topic due to its potentially broad application on both 2D images and 3D models used in different fields, ranging from computer-aided-design, medical imaging, computer graphics, to computational fluid dynamics. With recent advancement in both computer hardware and software, 3D models are commonly used in many applications because of their completeness of representation and the versatility. To deal with 3D objects, most techniques focus on shape matching or feature extraction, however, little work has addressed the problem of matching feature points on 3D models....[ Read more ]

Feature extraction and matching have become a popular research topic due to its potentially broad application on both 2D images and 3D models used in different fields, ranging from computer-aided-design, medical imaging, computer graphics, to computational fluid dynamics. With recent advancement in both computer hardware and software, 3D models are commonly used in many applications because of their completeness of representation and the versatility. To deal with 3D objects, most techniques focus on shape matching or feature extraction, however, little work has addressed the problem of matching feature points on 3D models.

This thesis presents an effective and stable template-based approach to match feature points on 3D deformable models, utilizing the surface invariant Gaussian curvature that characterizes the local surface shape. To support the implementation, a set of mesh models exported from POSER are used as both template and matching models.

In the template feature-extraction process, feature points representing the model landmark are extracted based on the curvature distribution, utilizing the combined techniques of curve-skeleton construction and sum of geodesic distance, which are two popular applications employed on 3D models. A new automatic skeleton construction method was developed.

Using the local negative Gaussian curvature and applying the feature-based template model, feature points can be matched onto the deformable models which share the same class of the models as the template. Experimental results indicated the proposed methodology shows good accuracy of feature point matching on different posture human models.