RaVioli: a GPU Supported High-Level Pseudo Real-time Video Processing Library

Abstract

Real-time video processing applications such as intruder detection system are now in demand and being developed. However, on general purpose computers, it is difficult to guarantee that enough CPU resources can be surely be provided. We have proposed a pseudo real-time video processing library RaVioli for solving this problem. RaVioli conceals two types of resolutions, frame rate and the number of pixels, from programmers. This makes video and image processing programmings more intuitive, but the performance may be lower by the abstraction overhead. To solve this problem, this paper proposes an improvement of RaVioli for supporting GPU platforms. For using GPUs effectively, a deep knowledge about them has been required, and this would have been a burden to programmers. The proposition on this paper provides an easy-to-use framework for developers. They can benefit from GPU without rewriting their RaVioli programs and get high performance video processing. The experiment results with image/video processing programs show that the proposed method improves the performance about 151-fold/164-fold in maximum against traditional RaVioli without rewriting programs, and about 30-fold/4-fold in maximum against a native C++ program.