GPU Optimization Of Video Feature Extraction Operators And Implementation Of Remote Computing

With the rapid development of information technology, Internet technology has been widely used. Web-based applications and services has become an important developing direction for software systems. In the meantime, the development of big data, artificial intelligence, and computer vision etc., are challenging the computing power. The general computing based on GPU are growing fast recently, e.g. CUDA(Compute Unified Device Architecture) launched by NVIDIA has greatly promoted the development of general purpose computing on the GPU, which has been widely recognized and used in fields such as biological engineering, financial calculations, image and video processing, fluid dynamics simulation and seismic analysis.This thesis proposed a web-based GPU remote computation platform, combining GPU’s parallel computing ability with the network technology, which allows users to conduct parallel computing tasks on the server-side GPU via a website. By using Internet technology, the high-efficient computing capability of GPU can be provided to users as a service. The GPU remote computing platform can provide online computing services for SIFT, STIP and MoSIFT feature extraction algorithms. In this thesis, STIP feature extraction is used as an example to demonstrate the CUDA parallel optimizing process, including memory optimization, reasonable division of threads and data, and rational use of CUDA streams.The experimental result on STIP feature extraction shows that the speed-up ratio of the parallel optimization varies with the video size and the number of feature points. The maximum speed-up ratio reaches 29 times. Compared with traditional local computing, this GPU remote computation platform enables user to access the GPU computing resources easily, no longer limited by times, spaces and facilities.