Automatic Player Indentification In Sports Video

Sports are very popular in the world. With the rapid development of the Internet and the huge amount of sports videos, sports video analysis become an interesting area with high commercial potential. However, the traditional manual annotation-based method could not meet the growing demand, therefore sports video analysis would be of great research value. For the problems of the current sports video analysis technology, this paper is mainly about player identification based on the jersey number recognition. The main contents are as follows:Firstly, player detection in the far lens. As there are always more than one entire player in the far lens, the HOG (Histogram of Oriented Gradient) feature which is commonly used in the pedestrian detection is adopted to detect the players. The rectangles with front-view, side-view and back-view players are selected as the positive examples, while the others without the players are as the negative examples. Then a SVM (Support Vector Machines) classifier is trained to determine whether the region contains player or not. In addition, considering that the upper body is more stable than the entire as players may make any actions, we perform two training in the experiment, one by the entire body, the other by the upper body. Experimental results show that the entire detection has a higher precision ratio while the upper detection has a higher recall ratio.Secondly, jersey region detection in the close-up shots. There are always players with the upper body in the close-up shots, and the letters and number on the jersey provide rich texture information. An algorithm based on the scene text detection is implemented to locate the jersey region. Besides, a method based on the face detection and the human proportion is innovatively proposed, as the face detection algorithm has been mature. The experiment results show that the latter has a higher accuracy.Thirdly, player’s team identification. Since the jersey colors of the opposite teams are contrasting, we extract the color feature to calculate the similarity and determine the player’s team. In this paper, the Hue cumulative histogram is employed as the color feature, which is shown has a higher accuracy compared to the RGB histogram and Hue statistical histogram in the experiment.Finally, the character segmentation and recognition. In the beginning, a preprocess with normalization and rotation corrected is applied. A KNN (K-Nearest Neighbor) classifier based affine transformation is implemented to recognize the letter and number. The experiment shows that this method has high accuracy and low computation.In summary, these research work provide novel idea to sports video analysis and has guiding significance.