Real time cloth modeling using parallel computing

Abstract

Cloth simulation technique has been successfully applied to the computer graphic and 3D apparel CAD filed for some years. However, although much effort had been devoted to improve the performance of the computational model, the computational requirement of the cloth simulation remains high and the running time is far away from the real time. Therefore in this dissertation, a new framework of the high-performance cloth simulation prototype system using parallel computing is developed. A simple mass spring model integrated with a set of force and velocity filter is used as cloth model. The dimension variation is monitored at every time step and the motion that cause the collisions are partially filtered. To improve the response speed of 3D simulation result for the 2D modification, a new methodology of reactive 2D/3D procedure to support the garment design modification is presented. Firstly, the designer initiates the modification by making adjustments on the 2D garment pattern boundary curve. The constraint forces required to deform the original curve to its new position is calculated. The triangular mesh of the original 2D pattern is deformed to fit the modified 2D pattern while the topology of the mesh remains invariant. Finally, the field of the equilibrium state parameter of the 2D pattern mesh is recalculated, based on which the corresponding 3D garment fitting simulation result is generated directly from the original simulation result. Another important advantage of the proposed 2D/3D reactive algorithm is that it well suits for parallel implementation. To alleviate the collision detection computing cost using parallel computing, the voxel based collision handling method is extended in this study. A multi level voxel based structure is built in the preprocessing. In this stage, the triangular faces on the human body mesh are computed to be associated with the corresponding voxels. Due to the nature properties of the voxel based method, the voxelization information can be rapidly updated during the simulation. Hence the time consuming elementary intersection checking is constrained within a small number of voxels. Furthermore at every time step, a face cluster hierarchy is dynamically built and able to improve the computation complexity of self collision detection algorithm from O(n 2)to O(n · m). Finally, this thesis exploits the parallelism of the algorithm of cloth simulation on the networked multi-computers. The parallel algorithms of cloth simulation and collision detection are presented and a dynamic load balance algorithm is proposed on distributed memory and heterogeneous parallel architecture. The proposed dynamic load balance algorithm takes both the cloth deformation computation and collision detection computation into account and can be regarded as a general scheme for the complex 3D scene simulation on the networked multi-computers, in which the simulated 3D objects interact with the other objects very often and hence both the spatial collision detection and simulation computation are critical computation tasks.