Complex scene modeling and segmentation with deformable simplex meshes

In this thesis we present a system for 3D reconstruction and segmentation of complex real world scenes. The input to the system is an unstructured cloud of 3D points. The output is a 3D model for each object in the scene. The system starts with a model that encloses the input point cloud. A deformation process is applied to the initial model so it gets close to the point cloud in terms of distance, geometry and topology. Once the deformation stops the model is analyzed to check if more than one object is present in the point cloud. If necessary a segmentation process splits the model into several parts that correspond to each object in the scene. Using this segmented model the point cloud is also segmented. Each resulting sub-cloud is treated as a new input to the system. If, after the deformation process, the model is not segmented a refinement process improves the objective and subjective quality of the model by concentrating vertices around high curvature areas. The simplex mesh reconstruction algorithm was modified and extended to suit our application. A novel segmentation algorithm was designed to be applied on the simplex mesh. We test the system with synthetic and real data obtained from single objects, simple. and complex scenes. In the case of the synthetic data different levels of noise are added to examine the performance of the system. The results show that the systems performs well for either of the three cases and also in the presence of low levels of noise.