Based On Multimodal Fusion Tennis Video Semantic Analysis And Research

Sports video retrieval is an important research direction in the field of content-based video semantic analysis. As a branch of study, tennis video analysis is becoming more and more important.In this thesis, using the games characteristics, we detect exciting events in the tennis video and construct a prototype system for event detection. This system achieves the video structured analysis, audio flow analysis and multi-modal fusion detection of exciting events.Video structured analysis is the key to the whole video retrieval because it directly affects the final detection of exciting events. This section mainly includes shot boundary detection, keyframes extraction, shot classification, and the slow motion replay detection. Shot boundary detection algorithm is based on the main color ratio difference and color histogram difference. This algorithm can not only obtain the frame difference threshold automatically, but also can improve the adaptability and flexibility. After studying the inherent features of tennis video, we propose a new detection algorithm which is based on the main color, tennis court line and the target object characteristics. This algorithm classifies lens into three types:On_Field lens, Close_Up lens and Off_Field lens. According to the logo lens which happens at the beginning and at the end of the slow motion, a new algorithm of logo lens detection is proposed which is based on the weighted shape invariant moments.A classification algorithm based on the continuous hidden markov model is designed. The core idea of the algorithm is firstly choosing the audio features such as the Zero-Crossing Rate, Short Time Energy, Mel-Frequency Cepstrum Coefficients and the Difference Mel-Frequency Cepstrum Coefficients, then using the Baum-Welch algorithm to train the parameters, and finally using Viterbi algorithm to classify the audio.Finally we summarize the algorithm of the existing events detection. Based on the In_Field lens, slow motion and the sound of shots,Net-approach event, Base-line Rally event and ACE event can be detected. Using the Visual C++6.0and Matlab7.1. we achieve an automatic tennis video analysis prototype system. Experiments have indicated that all these algorithms are effective.