Our investigations are aimed at elaborating a system for managing a digitized video data archive and to provide a collection of video data for easy access. These works are conducted in the context of the project for creating the electronic library "Kinoletopis' Rossii" (Cinema Chronicle of Russia) based on the collection of documentaries of the Russian State Film & Photo Document Archive.

The digitizing and storage of large amounts of visual material is no longer a problem from a technical point of view. The same is not yet so in case of searching for relevant visual information. Nowadays users of visual information systems refuse to be satisfied only with traditional query facilities on such items as title, author, subject or any other textual information associated with images.

Additional possibilities are required to formulate non-trivial queries on the content of video data, so a system must use various sources of information about the video material. Generally, video content can be conveyed in both narrative and image. The narrative content is commonly expressed by metadata (title, authors, date of creation, technical parameters, content description, etc.), speech and captions. Image properties of video are both attributes of low-level abstraction, i.e. simple computable visual features (like color histogram; color sets disposition in the image; texture and shape measures; optical flow velocities) and domain objects.

Generally, automatic scene understanding has not yet become an effective solution of the problem of access to an archive of digitized images, unless it is related to a restricted domain. Our project is aimed at managing documentary video material where the main character is generally a person (or a group of persons). That is why the most important complex object that the system must be able to recognize is a person. Our approach to object detection in video infers the extension of the detection methods are already known for static images and taking advantage of dynamic information not available for static images.

If a system is capable of detecting a person or, at least, a human face in consecutive frames in a video stream, the next thing to do is to connect the object recognized to a certain person that appears in the text description of the video material. The idea of automatic using text descriptions of video material content for this purpose and, generally, for understanding the visual content appears to be promising.