Study On Dense Stereo Image Matching Based On Parallel Computing

Stereo dense matching is very important in high resolution registration and 3D reconstruction. The dense matching method based on SIFT can obtain uniform and dense point cloud for object surface, which can help to obtain high resolution results, but it has low computing efficiency. Most of the 3D reconstruction and registration need to be real time, so accelerating the algorithm with parallel computing is meaningful. For designing parallel algorithm, how to choose appropriate technology according to the character of both the algorithm and device is a difficult problem. In recent years, GPGPU becomes a hot topic and has its own advantages and disadvantages compared with muti-thread. CUDA released by Nvidia in 2007 brought GPGPU into a wider field, but its performance and methods for optimization need to be researched farther.This thesis divides the stereo dense matching into several tasks and analyses the parallel and time consuming parts. Feature vector matching and dense matching parts are then designed into parallel programs with both CUDA and muti-thread. In the experiments, the performance in accelerating of these two technologies are compared and analyzed. For the CUDA based vector matching program, four different schemes are designed when task distribution for each thread is different. The performance of these schemes is compared in experiments. Also, the effect of several main optimization technologies in CUDA on the performance of program are tested in the experiments for CUDA based dense matching program.From the experiments, the computing efficiency of the parallel computing is far higher than signal computing. The technology and scheme need to be selected according to the specific algorithm and the practical hardware condition. The biggest difficulty for CUDA programming is the task distribution and the technology to optimize the communication between host and device or in device. This paper successes in accelerating the stereo matching algorithm to meet the requirements, provides theoretical reference for how to select appropriate technology when parallel program is designed for specific signal program. Also, it summarizes some useful experience for application of CUDA.