K. Sinha, A.D. Chowdhury, S.K. Ghosh, and S. Banerjee (India)
video surveillance, parallel system, cluster computing, on line video processing, load balancing, total completion time, average completion time, resource utilization.
Timeliness is an important issue for video based surveil lance and is often quantified by the delay between the time of availability of image frames from cameras and comple tion of their processing. Most existing commercial video surveillance systems focus on the issues of efficient storage and retrieval, remote monitoring, data streaming, forensics and limited real-time analysis - but not explicitly on the timeliness issues of large scale online analysis vis-a-vis re source utilization. In this paper we present a new load dis tribution strategy for on-line, large scale video data pro cessing clusters that are used as an aid to manual surveil lance. We propose a novel approach for fine grained load balancing, modeled as a minimization of average comple tion time problem. The proposed approach is robust in the sense that it is not dependent on the estimates of fu ture loads or on the worst case execution requirements of the video processing load. Simulation results with real-life video surveillance data establish that for a desired timeli ness in processing the data, our approach reduces the num ber of compute nodes by more than a factor of two, com pared to systems without the load migration heuristics.
Important Links:
Go Back
