The Grass-Soil Segmentation for Baseball Playfield Using Gaussian Mixture Model

• Main Authors: Yen-Hao Han, 韓顏壕
• Other Authors: Chung-Ming Kuo
• Format: Others
• Language: zh-TW
• Published: 2006
• Online Access: http://ndltd.ncl.edu.tw/handle/57668503530039158010

碩士 === 義守大學 === 資訊工程學系碩士班 === 94 === The purpose of this research is to segment the grass and soil regions of playfields in baseball videos. The playfield segmentation is very helpful to higher-level content analysis of sport videos such as scene classification, player tracking, and highlight detection, etc. In this thesis, we develop a new method of grass-soil playfield segmentation based on Gaussian Mixture Model. The Expectation Maximization (EM) algorithm is used to train the model parameters. In order to improve segmentation accuracy, we generate a particular GMM model for each baseball game. Each model is obtained either by selecting from a pre-trained model set, or automatic training directly from sample data. The proposed method is summarized as follows: 1. Compute the color feature maps of video frames, and extract the sample images automatically by using the feature maps and a priori knowledge. Then cluster the feature maps of the sample images by k-means algorithm. 2. Propose two model adaptation approaches to fit the color variations of different games: (a) establish offline a GMM model set which contains several models. One of the models is chosen to fit a particular baseball game. (b) train online a GMM model for each baseball game using training samples. 3. Correct the segmentation errors by horizontal scan filtering. 4. Detect the change of color distributions in background to update the GMM parameters.