PHOW Based Feature Detection For Head Pose Estimation

Face pose estimation is an important research of the field of computer vision the field of face recognition. It has extensive value in actual application and goodness in market prospect. Face pose estimation is a crucial step in face recognition system. At present, for the study of the face pose estimation at home and abroad have invested considerable human, material and financial resources. Especially in recent years, human has made a major breakthrough in the study of face pose estimation. According to the principle of algorithm implementation, the methods of face pose estimation can be divided into the following kinds: the methods based on template, the methods based on the detection, the methods based on facial geometric features, the methods based on the characteristics of the regression, the methods based on subspace learning, the methods based on local constraint model and so on. According to different types of data source, the methods can also be divided into the methods based on color facial gesture images, the methods based on depth images, and the methods based on 3D graphics images, etc. After analysis and comparison of the various face pose estimation methods, we found that in actual application of the above mentioned methods have some limitations.In this paper, we proposed face pose estimation algorithm based on PHOW. Among them, the SIFT scale-invariant feature transform algorithm, as a kind of local feature descriptor, has rotation, scaling and affine transformation invariance. For illumination changes, the change of observing perspective, partial occlusions and noise sensitivity is not high. Bag of Words model was first applied to text classification, after being introduced in the 2D images processing. BOW model help image semantic understanding, more blocked complex background change effect on the identification results. But because the model lose sight of the structure and location information of images, so we will use pyramid the scale space and location information to compensate for the missing part structure, in order to improve the effect of the face pose estimation. In this paper, the related experimental results show that the proposed for face pose estimation algorithm has good capability of identification and estimation.The main contribution and innovation of this paper lies in:Take face image SIFT local feature descriptor to describe the change of facial gestures, combined with the Bag of Words model, pyramid scale space and support vector machine(SVM) classifier to complete the estimate of the human face gesture. Experimental results show that the horizontal face posture rotation angle is large, high robustness, with perfect recognition rate.