Action Recognition in Multimedia Streams

It is well accepted that the rise in the proliferation of inexpensive digital media collection and manipulation devices has motivated the need to access this data by content rather than by keywords. The requirements of content based access are well understood by the digital media research community and there is no need to elaborate further here. Parsing multimedia streams by detection and classification of action implies modeling the dynamic nature of visual and audio features as they evolve in time. The Hidden Markov Model (HMM) has long been used to model dynamic behaviour in audio signals. Its power to capture complex behaviour in that domain has led to widespread use in visual content analysis because of the non-stationarity inherenet in those signals. However, subtleties in the application of HMMs are often unclear in the use of the framework in the visual processing community and the latter portion of this chapter sets out to expose some of these. Three applications are considered to motivate the discussions: actions in sports, observational psychology and illicit video content. Sports: Work in sports media analysis and understanding has been conducted for a decade now with clear motivation provided by the huge amount of sports media broadcasting on internet and digital television. An overview of content analysis for sports footage in general can be found in [22]. Action recognition here involves detection of certain plays and situations as dictated by the game domain e.g. pots, goals, wickets and aces. Illicit Content: The distribution of pornographic materials has also benefited from the digital revolution [6]. This kind of material is illegal in the workplace and is referred to as illicit content in this article. The issue of filtering this material has been of major concern since the introduction of the web in the early 1990’s. Pixalert’s ‘Auditor’ and ‘Monitor’, FutureSoft’s ‘DynaComm i:scan’ 2 and Hyperdyne Software’s ‘Snitch’ all provide image and