Research On Key Technology For Scene Analysis And 3D Reconstruction In Soccer Video

The explosive growth in the amount of multimedia information necessitates the development of content-based video indexing and retrieval techniques. Recently, sports video content analysis has become a hot research area. This research work's main objective is to automatically extract high-lights pieces from a video sequence for users quickly browsing and enriching them with different represent forms. Although great progress has been achieved in the past ten years, the successes of some systems can not satisfy users'requirement and is far from the research target.This paper focuses on some key technologies in soccer video analysis and 3D scene reconstruction, in order to make them more robust and applicable. The main work includes the following three points: 1) automatically detecting playfield. This technology is commonly used in soccer video analysis and 3D scene reconstruction; 2) moving objects detection and tracking. In sports videos these objects are ball and players; 3) 3D scene reconstruction based on monocular broadcast video. The main novelties consist of the following points:1. We propose a new algorithm for playfield color detection based on intrinsic image. As a key technology in sports video analysis, it is affected by shadow heavily. The proposed algorithm can overcome the problem. The algorithm first detects dominant color in the color histogram for a sequence. Since similar surface's intrinsic image has symmetrical distribution along the right projection direction, the camera sensor can be calibrated. As the projection direction directs, the algorithm justifies whether the second dominant is the playfield color by their intrinsic image distribution. Moreover, when the shadow regions and non-shadow regions are determined in playfield, the camera's sensor can be fined calibrated by maximize the similarity of the two regions'intrinsic image distribution. Experiment show that the algorithm can detect playfield colors even including shadow.2. We propose an adaptive playfield detection algorithm. Based on the detected playfield color, the algorithm exploits GMM to model these colors distribution, whose parameters are updated by incremental expectation...