Qualitative Camera Motion Classification for Content-Based Video Indexing

Due to the fact that the camera motion usually imply some hints which are helpful in bridging the gap between computationally available features and semantic interpretations, extensive researches have been executed to extract them for various purposes. However, these strategies fail to classify the camera rotation; furthermore, their performance might be significantly reduced by considerable noise or error in extracted features. In this paper, a robust camera motion classification strategy is proposed. We use the mutual relationship between motion vectors for motion classification. Given any two motion vectors in each P-frame, four types of mutual relationships between them are classified, then, a 14-bins feature vector is constructed to characterize the statistical motion information for the P-frame. Finally, the qualitative classification is executed by considering all achieved statistical information.