Abstract
Information retrieval from multimedia databases has become an urgent problem. Its solution can be facilitated by describing the content of multimedia databases using a variety of ways. In a video database, the options can be caption, speech, audio, image features etc. Presently, the MPEG-7 framework deals with standardisation of the multimedia content description techniques. Image features, such as motion, colour, texture and shape, are used for image annotation. The research described here is concerned with the annotation of sports video as part of an EU project ASS AVID. The framework of ASSAVID is similar to MPEG-7. The focus of the research is to develop motion feature descriptors. Motion description becomes increasingly attractive because motion features encapsulate temporal information. However, problems plaguing low-level motion processing impede the research on high-level motion analysis. This becomes more severe in applications with real-life video. In our research, human motion is adopted for sports annotation because sports involve a number of human behaviours. Human motion analysis has a wide spectrum of applications, such as surveillance, medical imaging and information retrieval. Yet there are no techniques directly related to this topic in MPEG-7. One of the useful descriptor of complex human motion is motion periodicity. However, among the existing techniques, only a few successful attempts at periodic motion description have been reported in real-life video. In this thesis, we present a novel method for sports video retrieval using periodic motion features. We focus on modelling human motion and this is accomplished by solving several sub-problems: A novel non-rigid foreground moving object detection algorithm is developed for complex real-life video. The algorithm is used to process low- level motion and segment out the human body from images with least computational expense. Innovative sport templates are constructed for human behaviour description using periodic motion features. They represent sport types in ASSAVID. Motion feature vectors are built using the templates. Motion feature classification is accomplished using a neural network. The proposed method has been tested on the ASSAVID database, which contains more than 800 minutes of real-life video from the BBC 1992 Barcelona Olympic Games. In total about 810,000 images have been processed to test motion features. Four types of different sports are tested. The experimental results show the proposed method to be successful.