Research On The Method Of Event Detection For Tennis Video

With the vigorous development of sports and multimedia technology,more and more people watch sports video through the Internet to pay attention to sports competitions.However,Internet video data grows rapidly.How to quickly locate their interested contents has become an urgent needs of users,video analysis and semantic event detection become the key to solve this problem.In view of the problem that tennis video is too long,and viewers want to find what they are interested in as soon as possible,we proposed a system to detect tennis video events based on fusion of multimodal information.Tennis video event detection mainly combines visual features,text information and domain knowledge.We use the method that combining the candidate segment selection with the singular value decomposition to reduce the eigenvector's dimension to detect shot boundary,it improved the accuracy and reduced the computational complexity.Analyzing the characteristics of the scoreboard in the tennis video,the scoreboard template is extracted according to the cumulative pixel difference and the edge detection method,and we divide the exact location of each information by image processing.Using the template matching technique,we divide the game video into scoreboard sequences and no scoreboard sequences,the scoreboard sequences are classified into global field shots(play shots)and non-global field shots(break shots)using the dominate color rate and the edge pixel rate,and we detect the replay shots from the non-scoreboard sequences.For play shots,we use the optical character recognition technology and the player name fuzzy match method to get the exact player's name.The score numbers are identified using the support vector machine classifier,and finally we get a complete score sequence.Analyze the structural characteristics of tennis video,we divide the complete video into video segments that point is the basic units.Combining the visual information,text information with tennis domain knowledge,we can carry out tennis video event detection,a total of 10 categories of semantic events were detected: game,tiebreak,deuce,advantage,break point,break,game point,rally,double fault,Ace.The experimental results show that the proposed event detection method plays a good effect on the detection of game,tiebreak,deuce,advantage,break point,break and game point.The detection of the rally,double fault and Ace needs to be further studied.