Implementation of Video Abstract Algorithm Based on CUDA

Abstract

The dynamic video abstract is an important part of video content analysis. Firstly, the objective of motion is analyzed, and the objective of the movement is extracted. Then, the moving trajectory of each target is analyzed, and different targets are spliced into a common background scene, and they are combined in some way. The algorithm uses Gaussian mixture model and particle filter to do a large number of calculations to achieve the background modeling and the detection of moving object. With the increase of image resolution, the computing increased significantly. To improve the real-time performance of the algorithm, a video abstract algorithm based on CUDA is proposed in this paper. Through the data analysis and parallel mining of the algorithm, time-consuming modules of the calculation, such as Histogram equalization, Gaussian mixture model, particle filter, were implemented in GPU by using massively parallel processing threads to improve the efficiency. The experimental results show that the algorithm can improve the calculation speed significantly in NVIDIA Tesla K20 and CUDA7.5.