Video structuring, indexing and retrieval based on global motion wavelet coefficients

This paper describes an approach for video structuring and indexing. It relies on motion wavelet coefficients directly estimated from image sequence. These coefficients provide a multi-scale characterization of optical flow. They allow one to define dominant and local motion descriptors, respectively related to camera and object displacements. We use dominant motion descriptors to perform a temporal segmentation of the sequence. Shots extracted are characterized in-term of dominant motion properties and indexed by using descriptors related to the local motion content. These operations allow one to retrieve shots, by example queries, according to only dynamic content of the scene and not camera displacements.