Motion descriptors for semantic video indexing

Content-based video indexing is a field of rising interest that has achieved significant progress in the recent years. However, it can be retrospectively observed that, while many powerful spatial descriptors have so far been developed, the potential of motion information for the extraction of semantic meaning has been largely left untapped. As part of our effort to automatically classify the archives of a local TV station, we developed a number of motion descriptors aimed at providing meaningful distinctions between different semantic classes. In this paper, we present two descriptors we have used in our past work, combined with a novel motion descriptor inspired by the highly successful Bag-Of-Features methods. We demonstrate the ability of such descriptors to boost classifier performance compared to the exclusive use of spatial features, and discuss the potential formation of even more efficient descriptors for video motion patterns.