Real-Time Event Detection From Broadcasted Baseball Video

• Main Authors: Ying-chung Zhu, 朱英菖
• Other Authors: Chaur-Heh Hsieh
• Format: Others
• Language: zh-TW
• Published: 2005
• Online Access: http://ndltd.ncl.edu.tw/handle/64743396282593524592

碩士 === 義守大學 === 資訊工程學系碩士班 === 93 === Content-based video retrieval has received wide attention recently. It is basically based on the low-level features of video signals. However, there exists a large gap between low-level features and high-level semantics. In recent years, many efforts in the literature have been made to bridge the gap because the semantic information will greatly facilitate users to access the large video libraries. This work focuses on the event detection from broadcasted baseball videos. The caption and visual information are extracted and then combined to effectively detect various events such as hit, homerun, double play and so on. First of all, the caption pattern in video is extracted and then employed to detect the play shot and non-play shot. Then, the non-play shots are removed and play shots are classified into four types of scenes using the extracted keyframe data of the shots. The classification is based on a trained scene codebook. The scene codebook consists of scene patterns. Each pattern is constructed by a 12-dimension color index. Finally, caption data and scene types are fused to develop an event detection rule. The proposed method is simple and effective. Simulation results indicate that our method achieves good detection accuracy and it operates in real time with Pentium xx computer. The event detection technique is very helpful for the design of sport video applications such as video highlight, summary and retrieving system.