Monocular Head Pose Estimation Based On Paraperspective Transformation

Abstract:The application of head pose estimation has become more and more prevalent in artificial intelligence, pattern recognize, intelligent human-computer interaction etc, and is of great significant either in human daily life or in safe production. This dissertation discusses the methods to estimate the head pose based on the computer monocular vision.A new method used to enhance the image, the lifting anti-symetrical biorthogonal wavelets transformation, is exploited. The method completes the multiscale decomposition of facial images, firstly. Secondly, to these decomposited informations, the reverse lifting wavelet transformation is performed. Then, the edges are detected by calculating the modulus maximum of this picture. Finally, the dates obtained by reverse lifting wavelet transformation are fused to enhance the image.Based on the CgCr color space feature edges detection and geometry informations, the feature location and corner detection algorithm is proffered. Firstly, eyebrow and eye rigions are roughly located. Secondly, the face region is accurately positioned according to the geometry relations between the eyes and the facial region. Thirdly, the edges of eye areas are detected out. And a new face region is reconstructed with the obtained informations to detect the nasal tip and mouth corners. Lastly, the facial plane model is founded with the nasal tip and mouth corners detected out.The method of head pose estimation based on paraperspective transformation correlation for the plane formed by facial feature points is selected as the calculation of head pose estimate. Firstly, the front-facial plane model is collected by the method which based on skin color segmentation with Adaboost front-facial classfier. The optimal front-facial plane model optimization algorithm is used to get the best front-facial plane, and current image plane model is combined to solve the transformation matrix.An un-embedding vehicle safety driving assistance system is designed on the basis of the analysis for the relationships between the fatigue and head pose in Pitch degrees-of-freedom, the distraction and the pose in Yaw degrees-of-freedom. The Yaw and Pitch angle threshold used to judge the states of driving are ascertained by geometric modelling and tests. Experimental results showing the technique can carry out the dangerous states of driving, and is meaningful to containing the occurrence of traffic accidents.