Tsung-Sheng Fu , Hua-Tsung Chen , Chien-Li Chou , Wen-Jiin Tsai , and Suh-Yin Lee Visual Communications and Image Processing (VCIP), 2011 IEEE, 6-9 Nov. 2011 Introduction System overview Camera calibration Player extraction and tracking Screen-strategy analysis Experimental results Conclusions Sports video analysis o Bring the audience efficient viewing of sports games • Highlight extraction and semantic event analysis [1, 2, 3]. o systems for tactics analysis and statistics compiling are in urgent demand [4, 5, 6] Basketball: one of the hottest sports o Chen et al. [7] proposed a 3D ball trajectory reconstruction algorithm which can be applied to shooting location estimation. o Chang et al. [8] introduced a wide-open warning system. To design a system capable of telling the executed tactics explicitly Scoring : the most important event, complicated task o Offensive tactics o Break the defense o Find open chance to shoot With the tactic information, audience can learn how plays are made, and professional coaches and players can analyze the offense tendencies and strategies. Screen:basic offensive tactic o Camera calibration o Player tracking video pre-processing: o Gathers reusable information o Accelerates the computation Content analysis: o Obtain their trajectories Geometric mapping between world coordinates and image coordinates. o Heavy load o Adapt the efficient court model tracking algorithm in [9] [9] D. Farin, S. Krabbe, P. H. N. de With, W. Effelsberg, “Robust Camera Calibration for Sport Videos Using Court Models,” in Proc. SPIE, pp. 80-91, 2004. Color filtering:detect white pixels Compute structure matrix within the pixel neighborhood: (b : Texture region width) Structure can be classified by evaluating the magnitude of the two eigenvalues. • 1 >> 2 : linear structure Hough transform Extract the longest horizontal and vertical lines by extracting the local maxima in the accumulator matrix Construct an accumulator matrix vote Camera parameters : homography matrix H. Time consuming Predicting the camera parameters for frame t + 1 based on the previously computed parameters for frames t - 1 and t. Computing the dominant color within the court region background subtraction k-means clustering Kalman filter o With the position predicted by the Kalman filter, we select the nearest candidate as measurement. o If a tracker is outside the court for consecutive n frames, it will be terminated o there are some candidates not tracked =>add new trackers Screen Detection o Two offensive players close to each other o At least one defender between the two offensive players standing close to each other Screen Classification o down-screen: screener moves to the baseline. o back-screen: the angle between the two directions is small, otherwise, mark as front-screen • moving direction of the screenee • the direction to the basket of the screenee Testing videos: Beijing 2008 Olympic Games: USA vs. AUS, ARG vs. USA, and USA vs.CHN with frame resolution of 640x352 (29.97 fps). total of 30 video clips o Randomly select 10 clips as our training data o Remaining 20 clips are testing data. We proposed a system that detects and classifies screens in basketball video Our proposed system is a one-pass scheme so that it can be applied to broadcast video. o The audience can learn offensive basketball tactics in real-time o Professional coaches and players can analyze the offense tendency of the opposing team efficiently.