Posture Estimation And Face Recognition Based On The3D Model

As one of the most important biometric techniques, face recognition technology has the characteristics of natural and non-contact. Compared with other biometrics, such as fingerprint recognition, iris recognition, it has obvious advantages. Due to the characteristics of non-invasive, face recognition systems should have the ability to identify facial feature in case of uncooperative or any other uncontrolled external conditions. The requirement to adapt to a variety of circumstances and conditions has brought a great deal of difficulty and challenge to face recognition technology, for example, the differences between the camera captured face images and the standard images in the face database could result in a substantial decline in recognition accuracy. Although many face recognition methods have achieved satisfactory results, but these effects are obtained in ideal circumstances, and still cannot meet the demand of a variety of actual circumstances.In the recent study of face recognition technology, the change of posture is considered one of the most important problems, which attracted the interest of many researchers in the field of pattern recognition and machine vision. Some promising face recognition methods have been proposed, such as tied factor analysis (TFA),3D morphable model (3DMM), eigen light-field (ELF) and illumination cone model, etc. However, because of their defects, these methods still cannot completely solve the problem of posture change in face recognition. With the current technology level, to get3D face data in a natural and low cost way is still considerably difficult, hence limited the application of the three-dimensional models in the face recognition system. In this paper, we reconstructed the3D face model by stereo vision and researched on posture changed face recognition by the depth data of the3D model. The main works of the paper are as follows:1. Discussed the principal of SIFT algorithm. In order to solve the problem of low efficiency caused by descriptors during the matching process, instead of the128-dimensions descriptor adopted in the original algorithm, a64-dimensions descriptor is put forward. Compared with the original one, the new descriptor expanded the statistical range of adjacent feature points, reinforced the feature information and reduced the feature descriptor’s dimensions. This method improved the matching rate. According to the experiment, results show that the matching time is greatly shortened.2. Researched posture changed face recognition method by the depth data of the3D model. By using the depth image of the face, the maximum curvature and the minimum curvature points of face image was calculated to serve as the tip point and the saddle point according to the principle of differential geometry and surface curvature. Based on the corresponding gray-scale images, it extracted the gradation characteristics and calculated left and right pupil positions. Then based on these feature points, this article calculated the face posture angle and achieved face pose estimation.3. Used the amplitude information of Gabor transform to do face recognition research. It counted the amplitude characteristics of face image, cascaded them into a feature vector of face, combined NLDA to do dimension reduction and facial feature classification, and realized face recognition eventually. This algorithm makes full use of the amplitude information of Gabor transform and reduces the dimension of face feature, both of which enhanced the robustness of the algorithm.This paper carry out research on3D reconstruction of face image, the depth value of the3D model is used to estimate the rotation angle of posture. Then rotate the3D model with the angle computed by depth image and project it into a2D image to compare with the standard face image in the library for face recognition.