GPU-based Vision System Of Design And Implementation Of Parallel Speedup

Computer vision technology has the characteristics of non-contact and high-automation. With the development of three-dimensional computer vision technology, using computer to simulate of human eye's visual function has become feasible. Since government now is starting to get aware of the importance of the research and development of internet of Things, The internet of Things based on computer vision has aroused high attention of all the academic community, industry filed and mass media. In this paper, the design and implement of the Three-dimensional measurement system is based on binocular stereo vision technology. However, because of the huge workload on system and the time assuming issue, it cannot meet the real-time requirement. Therefore, considering the powerful parallel computing capabilities of GPU, I designed a measurement and perception parallel processing algorithm based on GPU's characteristic, also made some improvements and optimizations for the algorithm, and finally increased the system operation speed, realized the real-time measurement and perception for non-contact object.Modern GPU (Graphics Processing Unit) uses relatively simple control logic and a large-scale general-purpose computing unit. With a powerful parallel computing capability and higher ratio of performance to price, GPU computing has become an indispensable part in IT filed. The use of GPU for general purpose computing has become increasingly common.AMD and NVIDIA have both launched their own programming model - Brook + and CUDA. But when comes to the transplantation of the program, AMD GPU and NVIDIA GPU are not compatible. It was not until Khronos Group released the open-computer language OpenCL, the language was able to fully work out the performance potentialities of heterogeneous platform and it was also transplantable.In this paper, the system has been tested on heterogeneous platforms which is composed of AMD phenom (tm) 2 X61055T 2.8GHZk and ATI HD5850. The test result shows that the use of GPU and multi-core CPU has an obvious effect on accelerating system speed; and the speedup rate can reach up to 3000 times, thus meet the real-time requirements of system.