3D Gaze Recovery in Large Environments Using Visual SLAM

This work describes a multi-component vision system that enables pervasive map-ping of human attention. The key contribution is that our methodology enables full 3D recovery of the gaze pointer, human view frustum and associated human cen-tered measurements directly into an automatically computed 3D model. We apply RGB-D SLAM and descriptor matching methodologies for the 3D modeling, lo-calization and fully automated annotation of ROIs (regions of interest) within the acquired 3D model. This methodology enables fully automated processing of hu-man attention, without artificial landmarks, in indoor natural environments.