Motion based situation recognition in group meetings

We present an unobtrusive vision based system for the recognition of situations in group meetings. The system uses a three-stage architecture, consisting of one video processing stage and two classification stages. The video processing stage detects motion in the videos and extracts up to 12 features from this data. The classification stage uses Hidden Markov Models to first identify the activity of every participant in the meeting and afterwards recognize the situation as a whole. The feature extraction uses position information of both hands and the face to extract motion features like speed, acceleration and motion frequency, as well as distance based features. We investigate the discriminative ability of these features and their applicability to the task of interaction recognition. A two-stage Hidden Markov Model classifier is applied to perform the recognition task. The developed system classifies the situation in 94% of all frames in our video test set correctly, where 3% of the test data is misclassified due to contradictory behavior of the participants. The results show that unimodal data can be sufficient to recognize complex situations.