A Study On The Parallel Computing Methods Of Visual Hull Based On CUDA

Three-dimensional reconstruction based on video images is an important research subject in computer vision, graphics, virtual reality and other fields, its goal is to recover three-dimensional scene information using one or multiple two-dimensional video images. The modeling method based on visual hull can reconstruct the corresponding three-dimensional model of an object according to the silhouettes of images acquired with multi-cameras in interactive rate, this method commonly used in three-dimensional modeling. But the visual hull method can only be done in real-time under certain conditions, in order to better apply it to virtual reality interactive and three-dimensional modeling systems, we need to optimize the efficiency of the algorithm, to ensure real-time performance of modeling. After analysing various modeling methods based on visual hull and the existing efficiency optimization strategies of visual hull, combined with GPU programmable mechanism and the meaning of parallel computing, This paper aimed at how to improve the speed of modeling, and focused on the research of efficiency optimization based on hardware acceleration of visual hull modeling method.1. First, we researched the idea of visual hull modeling, classification and main methods, analyze the existing efficiency optimization methods and research results, then combined with the architecture of modern graphics hardware (GPU), we discussed the parallel computing power and programmable mechanism of GPU, and ultimately determined the method how to apply GPU parallel computing to visual hull modeling.2. Through researching visual hull modeling method, combined with the thought of multi-threading and parallel computing, we studied parallel computing model of visual hull. We decomposed data pre-operation and generation algorithm of visual hull into subtasks in detail, and proposed the parallel model. On this basis, we provided a GPU-based simulation platform for computing visual hull, related researchers can do researches on algorithm optimization and simulation experients on this platform.3. Based on analysis of the existing GPU-based efficiency optimization methods of visual hull modeling, a parallel computing method of visual hull based on CUDA was proposed. This method takes advantage of parallel computing power of GPU to accelerate modeling, converting the generation algorithm of voxel-based visual hull into the process of parallel CUDA thread blocks. It makes full use of independence between the voxel data, parallel computing the states of each voxel and its vertexs. Meanwhile, In order to improve data transmission speed between main memory and graphics memory in the modeling process, this method also does some parallel optimization about data storage and transmission, and parallel execution between tasks is realized to further improve the modeling efficiency.Finally, the parallel algorithm above was realized on the simulation platform, with the traditional CPU-based serial algorithm of visual hull, we did comparison of experimental data and analysis of algorithm performance. The results shown that, under the premise of ensuring the accuracy of modeling, the parallel computing method of visual hull we proposed can quickly reconstruct complete visual hull model of target object, and the acceleration effect is obvious.