Team Intention Tracker — TInT

Researcher: Martin Gierisch

Runtime: 2006 -

Smart Environments need to identify user goals to enable the autonomous and proactive planning of appropriate assisting strategies. Recognizing the needs is addressed by Intention Analysis. This becomes a central challenge, especially if multiple users are observed exclusively by noisy heterogeneous sensors. We use a model driven approach based on dynamic Bayesian networks.

A small team meeting scenario defines the conditions for testdriving our approach: A team of three people is holding a meeting in our Smart Environment. They all like to present a short talk and therefore agreed on an agenda. But obviously, there is a slight probability that the team will change the agenda during the meeting.

The model for the described scenario is defined and implemented using Team Intention Tracker — TInT This suite provides the tools for team intention model definition and simulation of sensor data. It enables online and offline filtering (particle filters) of observed or simulated sensor measurements as well as learning of suitable statistics parameters using smoothing.

A socket driven visualization frontend provides online insights of the filtering process. The large visualization area on the left in the screenshot depicts the Smart Appliance Lab as 2D-map. It shows one possible room topology — adequate for a meeting scenario. Dark grey areas represent obstacles, e.g., walls or furniture. It can show the particle cloud for each team member as well as the most probable location estimates. The measured or simulated user positions are visualized by filled labeled circles.

The lower left corner of the figure depicts the probability distributions for the team intention and the right column holds the corresponding user intention probability distributions for team members A, B and C. The currently inferred team intention in this figure is B Presents. So, the user B is on his way to the presenting stage while user A walks back to take a seat to listen. Based on the team objective inferred, the room may automatically configure to support this goal (e.g., the current speaker's presentation is mapped to one of the displays and the lighting is adjusted).

Further TInT includes a bunch of lightweight evaluation tools that allow analysis and interpretation of intention recognition for recorded team meetings wrt. accuracy and delay of decision making or prediction horizon of specific models.

Key publications

Martin Giersich and Thomas Kirste. Effects of Agendas on Model-based Intention Inference of Cooperative Teams. In Proceedings of CollaborateCom2007 - The 3rd International Conference on Collaborative Computing: Networking, Applications and Worksharing, White Plains, NY, USA, November 12 - 15 2007. IEEE Xplore.

Martin Giersich and Thomas Kirste. Task Models for Inferring Team Intentions. to appear in KI, November 2007.

Martin Giersich, Peter Forbrig, Georg Fuchs, Thomas Kirste, Daniel Reichart, and Heidrun Schumann. Towards an Integrated Approach for Task Modeling and Human Behavior Recognition. In Julie A. Jacko, editor, Human-Computer Interaction, HCII 2007, volume 4550 of LNCS, pages 1109 - 1118, Heidelberg, July 22 - 27 2007. Springer Verlag.