The Research Of GPU-based Analytics On Large Scale Video

The unstructured large scale video has large volume and large amount of information,and the analytics on it is difficult and time-consuming.With the increasing demand for big data analysis,Hadoop and Spark as a large data processing framework centered on structured data have been proposed.The emergence of big data framework provides a new idea for the storage and analysis of large scale video data.At the same time,GPU parallel computing technology in the field of video and image processing has made a greater breakthrough,which can provide a strong support for the process of large scale video data analysis.In this paper,we will explore the effective combination of GPU parallel computing and video content analysis.In the video preprocessing stage,NVCUVID video decoding API is used to implement the pan-format video decoding library,which effectively reduces the high CPU usage.The NVENC video coding API is used to realize the pan-format video encode library,and the coding speed is accelerated.In the extraction process of video abstract frame,an algorithm based on mixed Gaussian background modeling algorithm and Frame Difference Method is proposed for the sea ice video frame,and GPU parallel computing is used to accelerate it.The algorithms above are verified by three experiments.In the stage of video image frame content analysis,a fast calibration and extraction algorithm for connectivity area is proposed for sea ice video.The algorithm is applied to the analysis of sea ice image with high density,which cannot be effectively solved by the existing connectivity region search algorithm.Aiming at the video preprocessing library based on GPU parallel programming,Spark’s call interface is designed by using Spark’s pipe action.The Spark calling interface using JNI is designed for the external frame content analysis library.Finally,a GPU-based big data analysis platform for large scale sea ice video is designed and implemented.The Spark task of sea ice parameter analysis is submitted by using videos on HDFS as the input data source.The results show that our big data analysis method based on GPU can provide a new idea for more efficient large scale video analysis.