Research Of Highlight Detecting Method In Soccer Video

Soccer video impacts on people's daily life, and people's concern for football video is much more reflected in the attention to highlights such as goal and penalty. Nevertheless it is a stuffy and inefficient job to find the highlights that people is interested in form mass sports video data. Therefore, it is necessary to find an effective method to detect the highlights in soccer video automatically.It discusses the issue of automatic highlight detecting in soccer video in detail in this paper. Existing methods either detect only the interesting scene or represent and detect highlights using low-layer features. Method based on Bayesian Network and Dynamic Bayesian Network using low-layer features, middle semantic and event achieves very good results. Highlight detecting method proposed in this paper takes two steps: Firstly it annotates original shots into semantic shots serial and then builds Hidden Markov Model to infer and detect highlights. When semantic annotating, based on exiting achievements in shot classification research, it classifies short shot further into player's close-up shot and off-field audience shot with texture feature of key frame image. Off-field audience shot has a distinct texture characteristic while player's close-up shot takes a smooth one, therefore it is easy to distinguish them with digital image processing technology such as edge detection and close operation. When reaching the event detecting, it constructs the HMM with semantic shots as evidence nodes and object events as hidden state nodes. According to the edit rule of soccer video and domain feature, it gets initial model parameters from plentiful training material, trains the model to adjust the parameters continuously using Baum-Welch algorithm till the establishment of ultimate model. At last, it uses Viterbi algorithm to infer and detect the highlights, that is, input the semantic shot serials into the detecting system and it is to get state nodes serials, accordingly it figures out whether there are some highlights or not.Experiment result presents a veracity ratio up to 98% of the classification method proposed in this paper, and higher precision and recall using HMM to detect the two highlights of goal and penalty.