Development of a two layer classification system for situations in meetings

Abstract

Studying the human behavior in meetings requires the recognition of the individual actions of all participants. Typical situations like discussion and presentation can be characterized using different behavioral cues. In order to deviate an interpretation of meeting situations the persons have to be detected and reliably tracked. The estimation of the face pose can be used to infer who is the current focus of attention, because people naturally turn their faces to the current speaker. In this work we present an approach to recognize individual activities based on processing videos of group meetings. Persons are detected using the selective running average approach for background subtraction. The Viola-Jones method is applied to locate the faces in the image plane. For each of the detected faces a feature vector is calculated using principal component analysis. Face pose estimation is achieved with distance based comparison in the feature space. For action recognition a set of global motion features is extracted for each participant. The resulting vector sequences are used for action recognition using hidden Markov models. The correct recognition rates of the developed system seem promising with rates for individual activities between 62% and 100%.