Recently, the Bayesian network has been applied for semantic analysis for soccer video games. Although existing works are intuitive to yield adequate results, they lack of scalability and robustness. Most of these works are mainly to extract features in pixel-domain by using image processing methods. However, their systems are limited to Low-definition video games as a direct consequence of slow computing speed. This paper introduces a flexible and scalable scheme to structure High-definition soccer videos by using the parameters embedded in MPEG codec, and a real time highlight scene detection system with Bayesian networks.