Continuous 3D recovery of human gaze using multisensory tracking

(Paletta et al., 2013) proposed a SLAM methodology using RGB-D information to reconstruct the 3D gaze within an automatically reconstructed 3D model with high precision. Their work described for the first time 3D gaze in indoor environments without the requirement of artificial landmarks in the field of view. In this work we significantly improve the coverage of service. The main challenge is to bridge gaps in the largely continuous but vision based 6 DOF pose reconstruction caused by effects of motion blur and insufficient coverage of texture with high spatial frequency. We achieve a continuous coverage of 3D gaze recovery, complementing the head worn eye tracking glasses (ETG) sensor with a high accuracy but still wearable accelerometer sensor which is being attached to the glasses frame. Using a particle filter approach (Durrant-Whyte & Bailey 2006; Gustaffson 2010) for multisensor based information fusion in the 6 DOF localization ??? using the camera sensor, an accelerometer, the floor plan provided by the 3D model and WLAN based positioning, we achieve a position accuracy of ?5cm (avg.) versus ground truth data, over a course of about 50 m attention service task, in contrast to ?20 cm using the vision based approach.