A Real-Time Head Pose Estimation Using Adaptive POSIT Based on Modified Supervised Descent Method
In this paper, we proposed a real-time head pose estimation algorithm by extending Pose from Orthography and Scaling with Iterations (POSIT) (named Adaptive POSIT) method and modifying the Supervised Descent Method (SDM). Specifically, we used the modifie
- PDF / 1,192,217 Bytes
- 12 Pages / 439.37 x 666.142 pts Page_size
- 49 Downloads / 176 Views
2
1 College of Computer and Information, Hefei University of Technology, Hefei, China [email protected] Department of Computer Science, Hong Kong Baptist University, Hong Kong, SAR, China
Abstract. In this paper, we proposed a real-time head pose estimation algorithm by extending Pose from Orthography and Scaling with Iterations (POSIT) (named Adaptive POSIT) method and modifying the Supervised Descent Method (SDM). Specifically, we used the modified SDM for facial landmarks detection and tracking, and adopted adaptive POSIT to estimate head pose. In the feature selection stage, we extracted different features in neighboring facial landmarks instead of a single feature. In the facial landmarks selection stage, we used partial facial landmarks instead of the whole facial landmarks. The experiments show that our method can track facial landmarks robustly with tolerance to certain illumination changes and partial occlusion, and improves the accuracy of head pose estimation. Keywords: Head pose estimation
SDM POSIT Facial landmarks
1 Introduction Head pose estimation, which determines the rotation angles involving three directions of face image in 3D space, has many applications including human-computer interaction, multi-view face recognition system [1], driving attention monitoring, and so forth. The rotation angle involves three directions including pitch (up and down), yaw (right and left), and roll (in-plane), as shown in Fig. 1. The methods of head pose estimation mainly include three categories: (1) facial appearance-based; (2) classification-based; (3) models-based. Facial appearance-based methods assume that there is a certain relationship between head pose and the features (such as gray, color, image gradient) extracted from face image, and the relationship can be established by a statistic method with a large number of samples. Extracting gray features in image space is a primitive method. However, the dimension of the gray features is very high and it requires huge computation cost. So, we generally replace image raw data with the features of lower dimensional space. © Springer International Publishing Switzerland 2016 D.-S. Huang et al. (Eds.): ICIC 2016, Part I, LNCS 9771, pp. 74–85, 2016. DOI: 10.1007/978-3-319-42291-6_8
A Real-Time Head Pose Estimation Using Adaptive POSIT
75
Fig. 1. Head pose in different angles. (1) The original image; (2–4) Subfigures show the images after the roll, pitch, and yaw rotations, respectively.
Classification-based methods mainly include manifold embedding methods and nonlinear regression methods. The former replaces face image with a high dimensional space vector and extract linear features from testing sample by mapping from the high dimensional space to its subspace. Then nearest neighbor classifier is used for classification. The latter learns the nonlinear functional mapping from the global features (such as Histogram of Oriented Gradient) extracted from face image to one or more pose directions. Model-based methods usually apply facial geometric relationship o
Data Loading...
