The Design And Implementation Of Video Abstract System Based On Multi GPU Acceleration

With the development of science and technology and safety awareness growing, more and more appear in the surveillance video around us, and there is a huge video data. When the need to view the surveillance video of the incident, the traditional method is gradually by the human eye inspection of video, finally found the sensitive video segment. But because of the limitation of human, this kind of method in waste a lot of manpower at the same time, there may be an important target. Video abstract technology through intelligent algorithms, automatic identification of sensitive objects in video, at the same time to remove the redundant information in the video, so that people can not lose the important information through browsing the video and minimal cost. In the process of video processing, there are a large number of parallel operations, these operations if by CPU,will make the video abstract time greatly reduced. So the performance optimization of data processing directly determines the practicability of the system, so the use of GPU is of great significance to the video abstract. In order to further improve the speedup, we choose to use multiple GPU combined acceleration.Through the programming method of GPU,we choose the design of GPU accelerated the use of CUDA programming framework. In order to obtain better acceleration effect, the paper analyses the characteristics of each module in the system, and finally determine the prospects for the target extraction and target fusion module for GPU rewriting. In the discussion of various methods with the prospect of target fusion, we choose the background subtraction method. In the target from a block of CUDA programming, the design and implementation of single GPU acceleration, GPU acceleration image segmentation and multi GPU multi-channel video acceleration scheme.There search in the complexity and time-consuming of each key step in the process of extraction of the target, find out the objective factors that hinder GPU acceleration, and trying to solve. For graphics switching multi GPU context leads to failure problem, we put forward a new method of multi thread coordination work to increase efficiency. As for the data between CPU and GPU to send each other and conflict of CPU decoding, we chose to use the method of labeling with buffer. In the analysis of the system is put forward afterthe ideal line, more fit to rewrite scheme of practical application, and the main process of the program design.Last recorded each mode under the running time of the system. By comparing the verified several different versions of the acceleration scheme for improving the performance of the system has about 20 times speedup and the late system needs to be improved in the direction of raises the prospect.