A Method For Multi-Target Tracking In Soccer Videos

With the development of science and technology and the improvement of people’s life quality, the requirement of image and video processing technology is more and more popular. Audiences and broadcasters’ requirement of analyzing and processing soccer videos is also urgent. The detection and tracking technology of players in soccer videos is the basis of realizing many high-level semantic analyses. As the complexity of the broadcast soccer videos, the detection and tracking of players is very difficult. So, the player detection, team identification and tracking are discussed in this paper.For player detection, the advantages and disadvantages of several common field extraction methods based on the dominant color segmentation are analyzed. Then, the improved field extraction algorithm adapting different illumination conditions is proposed. There still exist the stadium regions and some remaining field lines in the binary image, so some necessary procedures are carried on. On the basis, connected component extracting algorithm is used to detect the player regions combined with some features.For team identification, based on the result of player detection, the supervision method is used to discriminate different teams. That is to say, the models of players and referees are selected manually, and the similarity between the target and the models is compared to judge his team. In this paper, the 2D histogram of H and S channels is used to save the color information and the Bhattacharyya distance is the similarity metric.For player tracking, the algorithm based on Kalman filter and region-based matching is proposed. That is to say, an adaptive searching region is set according to the position of Kalman prediction, and correct matching results are obtained based on some matching strategies in the searching region. For the occlusion problem, the changes of players’ mass center and area in consecutive frames are used to judge the happening of occlusion. Besides, for the occlusion of players in the same team, it is processed according to different situations of the mass center’s mean value and variance in former N frames.Experiments show that the proposed field extraction algorithm eliminates most of the field lines. So, it helps to enhance the precision of player detection. Meantime, the improved tracking algorithm can process the occlusion, splitting, emerging and disappearing problem well.