Highlights Detection Of Soccer Video With HIDDEN MARKOV MODELS

Sports video retrieval is an important branch of content-based video retrievalfield, and is a challengeable direction. As the member of sports videos, soccermatch video has a great number of audiences. These audiences can be grouped,considering their purposes of watching match. The focus and the request ofretrieval are different as the group changed, and it is necessary to retrieve soccervideo in semantic layer.There are two points in semantic recognition of soccer events.1. Soccer video is well structured in some degree, however its structure isloose compared with news video. One example is that the highlights (such as goal,pause, etc) only happened occasionally, and the time when they will happen can'tbe predicted. Furthermore scenes related to these highlights will change moreaccording to changing of the match. Because of these reasons, it is not appropriateto predicted highlights in soccer video with rule-based methods.2. Existent semantic recognition methods mostly based on player/balltracking technology. But the complexity of light conditions and player interactionmake this type of methods inappropriate.On the other hand, soccer video is made up through some regular pattern. Itmeans that highlights in soccer video happened along with some typical elements(such as typical shot transition, typical sound, etc). Therefore these highlights canbe recognized along with the recognition of these typical elements.With analyzing the characters of soccer video, a Hidden Markov Model(HMM) based soccer video highlights system is presented in this paper.HMM is a statistic-based model, and is good with simulating, predicting ofstochastic data. HMM is broadly used in visual information retrieval field inrecently decade, such as gesture recognition, face recognition, handwritingrecognition, etc. but it has not been deeply researched on the applying of HMM insemantic recognition of videos, especially in semantic recognition of sports video.Work in this paper has some innovation as described below.First of all, the characters of soccer video are analyzed based on the domainknowledge of soccer match. Considering that there is strong color correlationbetween frames in soccer video, and the global motion of frames can deliverplenty of information, a new shot boundary detection method is presented in thispaper. The procedure include three steps:1. Adaptive detection of dominant region in frames. 2. Perform primitive segmentation by calculating ratio of dominant region. 3. Perform further segmentation by detecting difference of global motion. Experiment results show that this method is effective. Furthermore shotsdetected by this way can deliver much information that is useful to the processingfollowed. Having detected shot boundary, characters of frames, such as number ofplayer, player ratio, global motion, etc, are extracted. In this step, several newalgorithms are presented, such as Player in Field Detection Method (PFDM) andtheir effectiveness is proved by experiments. Considering drawback of key frames method, a Group of Histograms Method(GHM) is presented to represent shot. In this method, shot is represented by 5character vectors gotten through statistics of 5 characters mentioned above. After that, the type of Markov chain is confirmed. Relationship betweenelements in video (such as shot, frame) and elements in HMM (such as hiddenstate, observation symbol) are explained thereafter. Then a preliminary HMM hasbeen accomplished. To overcome insufficiency of training data, a relatively reliable measurementdeduced from Baum-Welch reestimation formulas is adopted in this paper. Bytraining samples in dataset, parameters in HMM are confirmed. Observation symbol is represented by a group of character vectors. Theobservation symbols belonging to same state will change much along with thedifference of match. In other words, the observation symbols belong to one infiniteset, and their probability distribution can't be determined in advance. So theoriginal method to get observation symbol probability distribution is improved.The observation symbol probability distribution is represented with the distancebetween observation symbol vectors and state vectors. If the distance is small, theobservation symbol is similar to the state, and the observation symbol is morelikely to appear in this state. The distance is represented by the sum of distancebetween these vectors. The state sequence can be estimated after that. Viterbi algorithm is used.When the probability of one state sequence exceeds a threshold, a corner kickevent is confirmed.