Section: Application Domains
Context Aware Video Acquisition
Video communication has long been seen as a potentially powerful tool for communications, teaching and collaborative work. Continued exponential decreases in the cost of communication and computation (for coding and compression) have eliminated the cost of bandwidth as an economic barrier for such technology. However, there is more to video communication than acquiring and transmitting an image. Video communications technology is generally found to be disruptive to the underlying task, and thus unusable. To avoid disruption, the video stream must be composed of the most appropriate targets, placed at an appropriate size and position in the image. Inappropriately composed video communications create distraction and ultimately degrades the ability to communicate and collaborate.
During a lecture or a collaborative work activity, the most appropriate targets, camera angle, and zoom and target position change continually. A human camera operator understands the interactions that are being filmed and adapts the camera angle and image composition accordingly. However, such human expertise is costly. The lack of an automatic video composition and camera control technology is the current fundamental obstacle to the widespread use of video communications for communication, teaching and collaborative work. One of the goals of project PRIMA is to create a technology that overcomes this obstacle.
To provide a useful service for a communications, teaching and collaborative work, a video composition system must adapt the video composition to events in the scene. In common terms, we say that the system must be "aware of context". Computationally, such a technology requires that the video composition be determined by a model of the activity that is being observed. As a first approach, we propose to hand-craft such models as finite networks of states, where each state corresponds to a situation in the scene to be filmed and specifies a camera placement, camera target, image placement and zoom.
A finite state approach is feasible in cases where human behavior follows an established stereotypical "script". A lecture or class room presentation provides an example of such a case. Lecturers and audiences share a common stereotype about the context of a lecture. Successful video communications require structuring the actions and interactions of actors to a great extent. We recognize that there will always be some number of unpredictable cases where humans deviate from the script. However, the number of such cases should be sufficiently limited so as limit the disruption. Ultimately, we plan to investigate automatic techniques for "learning" new situations.
This system described above is based on an approach to context aware systems presented at UBICOMP in September 2002  . The behavior of this system is specified as a situation graph that is automatically compiled into rules for a Java based supervisory process. The design process for compiling a situation graph into a rule based for the federation supervisors has been developed and refined within the last two years.
Since 2004, we have demonstrated a number of systems based on this model. In the FAME project, we demonstrated a context aware video acquisition system at the Barcelona Forum of Cultures during two weeks in July 2004. This system was also demonstrated publicly at "Fete de la science" in Grenoble in October 2004, and exhibited at the IST Conference in Den Haag in November 2004. A variation of this system has been integrated into the ContAct context aware presentation composition system developed with XRCE (Xerox European Research Centre), and is at the heart of the CHIL Collaborative Workspace Service used in the IP Project CHIL. A context aware interpretation system for video surveillance is currently under development for the IST project CAVIAR.
This system is under permanent installation in a meeting room and in an amphitheater of the LIG laboratory. This installation is a part of the LIG plateform demonstrators.