The Twenty-First APT Standardization Program Forum
11 - 15 March 2013, Bangkok, Thailand / Document:
ASTAP21/INP-37
5 March 2013
ETRI, Republic of Korea
Framework for a 3-d based head pose estimation system
1. Introduction
This contribution provides the framework for three dimensional head pose estimation system
2. Framework for a 3-D based Head Pose Estimation System
Calculating the 3-D head pose is a fundamental process for unconstrained automatic face recognition systems. The 3-D head pose describes not only direction but also basic information such as the size and position of the head. Accurate information of the head pose is also essential for face recognition. In the case of 2-D face recognition the variation of the head pose results in a different recognition rate. Also, in the case of 3-D face recognition, it indicates the position in a 3-D space. Therefore, if the information is not accurate, it can result in a low recognition rate.
Basically, estimating the head pose requires calculation of the translation and rotation information in 3-D space. There are two main categories in existing methods: using an actual 3D head model or using an approximate head model. In this proposal, the 3-D model based head pose estimation will be discussed in the specific case. Figure 1 shows the example of 3-D head pose estimation.
Fig. 1. 3-D head pose estimation
Before defining the framework for 3-D head pose estimation system. We should describe the environment where the framework should be applied, first, in the case of using a single camera and in the case that there is a 3-D model which can be fitted to 2-D image. In these conditions, our framework should be working. Figure 2 shows the architecture of the proposed 3-D head pose estimation system. The system mainly consists of six modules; facial feature extraction, feature compensation, 3-D model creation, pose estimation, 3-D model updating, and posecompensation.
System Configuration
- Facial feature extraction
In this module, it is assumed that face is detected by a face detector.For the extracted face region, the facial features are extracted.
- Feature compensation
In video, the positions of the extracted facial features can be varied between frames. Therefore, the robust tracking method should be needed such as Kalman filter for stable feature tracking. Besides that, it can remove the high frequency noise.
- Pose estimation
The 3-D head pose is estimated by fitting a 2-D face image and a 3-D face model. The optimization algorithm can be used for minimizing the fitting error between them.
- 3-D model creation
In off-line, 3-D face model should be constructed. Depending on the quality of aninput image, the size and the shape of the 3-D model should be pre-calculated.
- 3-D model updating
User interface should be provided for correcting the accumulation error. During the facial feature and pose tracking, the error can occur.
- Pose compensation
After pose estimation, pose compensation process is needed. The result of the calculated pose can be affected for initial value or intermediate noise. Therefore, the proof module should be applied to correct the head pose.
Fig. 2. The framework of proposed 3-D head pose estimation system
As mentioned, there modules are basic function for estimating 3-D head pose. It will be used for constructing the basic structure of head pose estimation system.
3. Summary
This contribution haspresentedbrief introduction of the framework of 3-D head pose estimation system. Besides that, we describedthe minimum system components which should be equipped in the system.
References
[1]W. Zhao, R. Chellappa, P. J. Phillips, and A. Rosenfeld, Face recognition: A literature survey, ACM Computer Surveys, 35(4), pp. 399-458, 2003.
[2]E. Murphy-Chutorian, M. M. Trivedi, Head pose estimation in computer vision: A survey, IEEE Trans. Pattern Analysis and Machine Intelligence, 31(4), pp. 607-626.
ASTAP21/INP-371