A parallel architecture for video processing

Abstract

Video data consists of a sequence of frames that is produced at a constant rate and many applications in real-time require the processing of these frames executing compute intensive algorithms. To handle many of such applications in real-time, we developed a new architecture based on parallel processing. The parallel architecture for video processing has been developed at Queen Mary and Westfield College as a part of an European Union RACE II project called MONALISA. A multiprocessing kernel and a high level software environment called SAPS (self adapting parallel server) model has been developed for this architecture. This environment makes it possible to introduce a number of load balancing and data decomposition schemes that can be realised automatically in realtime by the kernel without any explicit inputs from the user. The developed architecture aims at applications such as image analysis algorithms, camera tracking, mixing captured foreground images and synthetically generated background images using depth values in real-time for Virtual Studios. In this paper we focus on the software architecture, frame buffer management and frame buffer access protocols. The system architecture and hardware is explained first. The standard frame buffer access protocol, SFBA, and the dedicated frame buffer access protocol, DFBA either of which address different needs of video processing are introduced. System performance evaluation, benchmark results and an analysis of DFBA protocol are given in detail.