A best view selection in meetings through attention analysis using a multi-camera network

Human activity analysis is an essential task in ambient intelligence and computer vision. The main focus lies in the automatic analysis of ongoing activities from a multi-camera network. One possible application is meeting analysis which explores the dynamics in meetings using low-level data and inferring high-level activities. However, the detection of such activities is still very challenging due to the often corrupted or imprecise low-level data. In this paper, we present an approach to understand the dynamics in meetings using a multi-camera network, consisting of fixed ambient and portable close-up cameras. As a particular application we are aiming to find the most informative video stream, for example as a representative view for a remote participant. Our contribution is threefold: at first, we estimate the extrinsic parameters of the portable close-up cameras based on head positions. Secondly, we find common overlapping areas based on the consensus of people's orientation. And thirdly, the most informative view for a remote participant is estimated using common overlapping areas. We evaluated our proposed approach and compared it to a motion estimation method. Experimental results show that we can reach an accuracy of 74% compared to manually selected views.