INTERACT: Inspection of Normal and Typical Encounters Requiring Asymmetric Collection and Tracking

Because social interactions are ubiquitous for both the police and the military, it is crucial to improve their outcomes. However, assessing the success or failure of social interactions can be rather cumbersome. In order to capture and measure these interactions, video, sound, movement, and other forms of data must be collected and analyzed. However these methods require that comprehensive data is collected from all individuals involved. What is needed are ways to"fill in the gaps"when data are missing. In this proposal, titled INTERACT, we propose to develop these methods using supervised learning through support vector machines and temporal learning through a Hidden Markov Model (HMM) representation. Supervised learning allows the system to predict missing data based on patterns in the available data. The Hidden Markov Models will then assess and predict the interaction dynamics. Linking these methods in a feedback loop will allow each learning method to benefit from the conclusions of the other. This methodology will be verified by assessing and predicting interactions within existing data sets.