Real-Time Head Pose Estimation Framework for Mobile Devices

Abstract

The head pose estimation technique predicts the rotation of the human head by analyzing a person's face in a digital image. The head pose estimation framework uses two processes for the estimation. The first step is the detection of the face and facial features using a Haar-like feature detector. Methods proposed in previous studies generally provided a low overall detection ratio of each facial feature. Therefore, the pre-processing step for storing the facial features as a template could be time consuming. We propose a calibration method that finds one eye feature that cannot be found on the front part of the face. The method was evaluated by conducting an experiment to measure the detection accuracy of the face and facial features. The second process is used for the template-matching algorithm while the facial features are being tracked. As the experiment proceeded, we measured the time required to execute the estimation on an Android device. The head pose estimation procedure uses the coordinates of facial features. The algorithms used in the proposed systems show that the detection and tracking processes require approximately 230 ms and 20 ms, respectively. In addition, the calibration method proved to be effective in terms of decreasing the detection failure rate by approximately 8 %. Thus, this result confirms the effectiveness of our method on mobile devices.