Automatic feature extraction from tennis videos for content based retrieval

This thesis presents an approach for automatic feature extraction from tennis videos for content based retrieval. Feature extraction for tennis videos involves locating and tracking players, detecting court lines, tracking the ball and shot detection. For detecting court lines, a Hough transformed based algorithm is presented. Background subtraction combined with shadow detection and removal is employed to determine the players' location accurately. A Minimum Boundary Rectangle is drawn to encompass the players in the segmented image to locate the player. Audio information extracted from the video is used to identify frames in which players strike the ball to reduce the search area for the ball to a window around the player. These frames are further analyzed to classify the type of shot as either a forehand shot or as a backhand shot. Results obtained on real tennis videos are presented to demonstrate the validity of the approach.