|Authors||V. S. W. Eide, O. Granmo, F. Eliassen and J. A. Michaelsen|
|Title||Real-Time Video Content Analysis : QoS-Aware Application Composition and Parallel Processing|
|Afilliation||Communication Systems, Communication Systems|
|Publication Type||Journal Article|
|Year of Publication||2006|
|Journal||ACM Transactions on Multimedia Computing, Communications, and Applications|
Real-Time content-based access to live video data requires content analysis applications that are able to process video streams in real-time and with an acceptable error rate. Statements such as this express quality of service (QoS) requirements. In general, control of the QoS provided can be achieved by sacrificing application quality in one QoS dimension for better quality in another, or by controlling the allocation of processing resources to the application. However, controlling QoS in video content analysis is particularly difficult, not only because main QoS dimensions like accuracy are nonadditive, but also because both the communication- and the processing-resource requirements are challenging.This article presents techniques for QoS-aware composition of applications for real-time video content analysis, based on dynamic Bayesian networks. The aim of QoS-aware composition is to determine application deployment configurations which satisfy a given set of QoS requirements. Our approach consists of: (1) an algorithm for QoS-aware selection of configurations of feature extractor and classification algorithms which balances requirements for timeliness and accuracy against available processing resources, (2) a distributed content-based publish/subscribe system which provides application scalability at multiple logical levels of distribution, and (3) scalable solutions for video streaming, filtering/transformation, feature extraction, and classification.We evaluate our approach based on experiments with an implementation of a real-time motion vector based object-tracking application. The evaluation shows that the application largely behaves as expected when resource availability and selections of configurations of feature extractor and classification algorithms vary. The evaluation also shows that increasing QoS requirements can be met by allocating additional CPUs for parallel processing, with only minor overhead.