Research On Streaming Video Recognition Method Based On Discrete GMM-HMM

With the development of broadband technology, streaming media technology has been morewidely used, including video transmission which is one of the main forms of communication in thestreaming media technology. However, due to the high bandwidth of streaming video technologyand the limited network resources, it is necessary to manage variety of streaming video trafficreasonably and effectively in network, and one of the most important task is multimedia trafficclassification and identification.This paper presents a method of streaming video traffic classification which is based on thediscrete Gaussian Mixture Model-Hidden Markov Model(GMM-HMM) model. Three commontypes of streaming video traffic are studied in this paper, including Full High Definition(Full HD)video files--HD video files with a resolution of1920*1088, Common Intermediate Format(CIF)video files--ordinary video files with a resolution of352*288, Quarter Common IntermediateFormat(QCIF) video file--video files which applicable to the mobile terminal with a resolution of176*144. This thesis focuses on the development of a discrete GMM-HMM structure which takesinto account the Inter-Packet Times(IPT) sequences. In this algorithm, each packet of the videostream sequence has a corresponding HMM underlying hidden state sequence. Besides, the first fewpackets of unknown data stream can directly show its category, leading to a faster classification.Discrete GMM is used to build HMM and the continuous observation sequence is divided byinterval discretization, thereby reducing the complexity of the recognition algorithm. In order toguarantee the accuracy of classification in different network environments, the author defines asimilarity measure between different HMM recognition model and develops an algorithm forcalculating the number of states in HMM recognition model. Enough states can ensure the accuracyof recognition under different network environments but the recognition speed will slow down ifthere are too many states. Therefore, a suitable number of states must be adjusted in HMMrecognition model, which can achieve the balance between accuracy and efficiency.