Slice representation of range data for head pose estimation

摘要

Visual estimation of head pose is desirable for computer vision applications such as face recognition, human computer interaction, and affective computing. However, accurate estimation of head pose in uncontrolled environment is still a grand challenge. This paper proposes a novel feature representation model for accurate pose estimation. In this model, a range image is divided into a set of simple slices that contain abundant geometric cues can be used to accurately describe the poses of a subject. This model provides a general framework for designing new features for head pose estimation. According to this model, design of a new feature model for describing a slice, then a new set of features is generated by combining all slices for describing range images. Due to the huge number of slices that can be generated from single range image, even a simple description model of slice can achieve robust performance. With the guide of this model, two novel range image representation models, which are Local Slice Depth (LSD) and Local Slice Orientation (LSO), are designed. LSD can be used for coarse estimation of head poses, while LSO can achieve accurate results. Moreover, in order to evaluate the performance of proposed representation model, an automatic head pose estimation method is implemented using a Kinect sensor. Firstly both color and range images captured by a Kinect sensor are used to localize and segment the facial region from background. Secondly, two novel integral images, namely slice depth integral image and slice coordinates integral image, are proposed to achieve real-time feature extraction. Finally, random forests are used to learn a stable relationship between slice feature descriptors and head pose parameters. Experiments on both low-quality depth data set Biwi and high-quality depth data set ETH demonstrate state-of-the-art performance of our method.