The Player Detection And Segmentation In Sports Video

Sports video is a kind of popular video media data. Content-based sports videoanalysis has become a hot research topic among lots of domestic and foreign scholarsbecause of billons of sports enthusiasts and extensive domain knowledge. Widespreadresearch on sports video analysis has been made. In earlier time, sports video analysisfocused on structure analysis of video content. Then recent years, sports video analysis hasbeen focusing on high-level semantic detection. Aimed to these high-level semanticresearches, the basic work is the players’ accurate detection and segmentation in the sportsvideo.There are a lot of characteristics and domain knowledge in sports video which can begreat useful for the high-level semantic analysis. The player detection and segmentationproblem is solved through a kind of object detection and segmentation method. As playersin different shot type with different size and other features, different detectors based onmid-level feature patch are trained. There are bounding box and polygonal area resultsrepresenting the players’ detection result, and with the help of superpixels segmentation,the superpixels in the polygonal can be labeled to be the players’ superpixel or thebackgrounds’ based on the relationship between the polygonal area result and the players’contour. The player segmentation results come out with the players’ bounding box and thesuperpixels’ labels treated as interactive information and inputted into the Grab Cutsegmentation algorithm.Experimental results show that the detection precision rate is rather high with thedetector based on mid-level feature patch. And at the same time, compared to the detectortrained with low-level features, the detection result produced by detector based onmid-level feature patch can provide a bit more priori knowledge for the later segmentation.But it’s difficult to make mid-feature patch to be more effective which should be focusedon and studied in the next step. In terms of the player segmentation, combing withsuperpixel labels and Grab Cut segmentation algorithm could greatly simplify thecomputation. Otherwise the players’ hair and arms are always cut off which result in thatsegmentation method should be improved for better segmentation.