Head Posture Estimation by Deep Learning Using 3-D Point Cloud Data from a Depth Sensor

Abstract

Head posture estimation is performed by capturing characteristic areas of the face, such as the eyes and nose, in images acquired from a camera installed in front of the subject. However, with this method, parts of the eyes and nose are hidden when the subject turns away and faces the side, making estimation difficult. In this letter, we aim to realize a more effective head estimation method than previous research using 3-D point cloud data from a depth sensor. We pursued the estimation of five head posture classes. In the proposed method, first, the 3-D point cloud data of the postures in the five classes are learned by a deep learning model. Next, the posture of the head is estimated using the model. In this letter, many verification experiments confirmed that the proposed method is very effective for head posture estimation with five posture classes.