Métodos GPGPU para simulación y visualización de modelos volumétricos interactivos

Abstract

In this thesis we address di erent computationally demanding problems in the elds of simulation and visualization, exploring novel GPGPU methods to provide solutions for interactive applications with a reduced time budget. In the rst part of this work, we tackle di erent limitations of the ChainMail algorithm for deforming volumetric medical models. We propose a novel parallel ChainMail algorithm for the e cient handling of large heterogeneous models, also accounting for interactive topology changes. The use of a novel scheduling method for the GPU addresses the sparse nature of the computation, allowing the use of much larger models in interactive applications. The second part is devoted to the online motion planning of soft robots addressed as a computational simulation problem. Speci cally, we propose an accurate and e cient model for hydraulic actuators and its usage in an inverse kinematics framework for the interactive control of actuated soft robots. An e cient estimation of the uid weight is provided by a novel parallel algorithm exploiting modern GPU capabilities. The last part of this work addresses problems related to the visualization of medical datasets through direct volume rendering. First, we propose a novel resampling algorithm for the interactive rendering of deformable medical models, capable of interactively providing a regular grid representation of the deformed dataset as input for the standard volume rendering pipeline. Next, we present a new tool for the the selection and visualization of regions of interest avoiding any kind of parameter tuning, aiming to eliminate trial-and-error approaches typically required for this task.