Research On Key Technologies For Typical Events Detection In Soccer Video

Recent years, with development of network and hardware, people are flooded with information. Therefore, content-based video retrieval became a hot topic, as a part of this, sports video retrieval is being researched widely. Soccer video is enjoyed by many fans, and plays an important role in daily life. We choose soccer video as our research object, and analyze and research the key technologies in typical events detection for soccer video, including shot boundary detection and low-level characters extraction (field and forbidden zone detection, player and goal detection), then combine the domain knowledge of soccer and use HHMM to infer shooting events and fouling events.As the first step of video analysis, shot boundary detection has been concerned by learners, and amount of methods have been proposed at previous years. According to specific details of soccer video and the existing methods, two-phase video shot boundary detection approach with various steps is proposed in this paper.Shot classification is very important to event inference, this paper proposed shot classification based on color and edge and classified shots into global view, zoomed-in shot, audience shot and close-up. In the light of analysis of soccer video, generally, goal will appears in shooting events, therefore, this paper proposed method of goal detection using Bayes classifier to classify goals. Moreover, slow-motion replay usually occurs after shooting or fouling, so this paper compared the relevant methods of slow-motion detection and chose a proper method in our experiment. For the field and forbidden zone and player detection, we study the existing methods and improve some of them according our actual situation.HMM is a statistic-based model, and is good for simulating, predicting of stochastic data and fit for event detection of soccer videos. However, we must construct HMM models for each event, and can not process video streams containing more than one event. HHMM is extension of HMM and can process temporal segmenting, namely, can segment and recognize more than one event. In this paper we make use of HHMM to model soccer video and detect shooting and fouling events.The experiments showed that our methods can achieve good results in shot boundary detection, shot classification, goal detection and detection of typical events.