Scene-Level Analysis for Tennis Sports Video using Weighted Linear Combination of Visual Cues

J. Han, W. Lao, and P.H.N. de With (The Netherlands)


Semantic analysis, visual cues, real-world court model, and event detection.


We present a scene-level analysis system that takes TV broadcasted tennis footage as input and produces a be havior analysis of the moving-players in the scene. To achieve this functionality, our algorithm relies on two mod ular blocks. The first one detects and tracks a number of key objects in the image domain, like moving-players and the playing-field. Afterwards, a camera-calibration algo rithm is applied that uses the lines of the court as a ref erence and transforms image coordinates to physical posi tions to compute the camera parameters. The second block firstly models several important events of a tennis game, such as service and net-approach, based on four real-world visual features provided by the first block. This paper pro poses a new improved model, since it weights the impor tance of each visual cue to different events, rather than pro viding a simple combination of these four visual cues as was done in previous work. Based on the new model, we can accurately determine what kind of event the current in put frame belongs to. Furthermore, we detect the start time and end time of each event using a simple but efficient tem poral filter. Our proposed system is capable of classifying each tennis play into three semantic categories, which are popular and familiar to most viewers. This paper presents details of the system, together with results on a number of tennis video sequences.

Important Links:

Go Back