Head-Pose-Based Attention Recognition on Large Public Displays

Estimating peoples´ attention to regions of large public displays has been a problem since those displays´ advent. Unlike with traditional interaction with a Web browser, you can´t calculate a clickstream. A method for estimating where users are looking could partly overcome this issue. Researchers evaluated how several factors (head movement, individual users, the users´ locations, and the amount of training data) affected the accuracy of attention recognition based on only the head pose. The results revealed three things. First, head movement in both the yaw and pitch directions insignificantly decreased the accuracy, compared to limited vertical or horizontal movement. Second, differences in accuracy of up to 16 percent suggest that you should train such systems on individual persons to achieve optimum recognition performance. Finally, calibration on multiple positions didn´t significantly enhance recognition, compared to training on a single position.