Research On Automatic Detection Of Soccer Game Events In Videos

Soccer game events detection is an interesting research topic in computer vision,and the algorithms that are developed during the studying of the games may have valuable helps for different purposes,such as for clip retrieval(posterity logging),for video synopsis(live logging),for real time transcoding for smart phones services,even for a better understanding of people's behaviors in other cases.In this thesis,automatic detection of twelve events of soccer game in videos is studied,and our research focuses on detecting cross-time(period-based)events and hierarchically connected events,besides,some basic problems,such as occlusions,ball detection,tracking,etc.,are also solved.The twelve events that are going to be detected are the main ones in a soccer game and they are kickoff,direct free kick,offense,penetration,formal defense,multiple layer defense,attack,approach,dribbling,dodging,smart positioning,and goals.The algorithms used in this study include visual feature extraction and scene classification to segment frames in order to decrease the number of frames to be processed.During the classification,the maximum exploitation of playfield locatives is considered and the hidden Markov model is used.By considering the quite different sizes of a player and the ball,we use a collaborative algorithm for players tracking,while the three other integrative trackers for ball tracking,which are the Kalman filter,the feature tracker and a location guessing tracker.Finally,events detection and classification are performed based on the imported information from the tracking stages about the players and the ball,where the velocities,the distances and the directions have been taken into account.The algorithms we proposed have been tested on 41 videos which include several full matches and some highlights(match summary),and the algorithms can effectively detect the twelve events listed above.Moreover,it is possible to detect more similar events in the future if our algorithms are extended.