Research Of Head Pose Estimation Method Based On Multi-feature Fusion

With the rapid development of artificial intelligence and biometric recognition technology,various information contained in the head pose changes can be widely used in fatigue driving,auxiliary clinical diagnosis,human behavior analysis and other fields in life,showing great research value and application value.However,in practical application,many uncertainties will affect the estimated results of head pose,such as changes in the intensity of the illumination,exaggerated facial movements and so on.Therefore,this paper mainly studies the head pose estimation method.This paper works on a head pose estimation method based on multi-feature fusion.The details are as follows:1.The face detection method of multi-pose is studied.An improved multi-pose face detection method based on Adaboost is adopted.Firstly,on the basis of Adaboost algorithm,an improved cascade classifier structure is used for multi-pose face detection to improve the face detection rate under multi-pose.Then,based on the results,the false detection rate can be effectively reduced by establishing skin color model and screening face.Compared with similar face detection methods,this method is more precise for recognizing multi-pose face.2.Feature extraction is also on the studying list.In this paper,the method adopts multi-feature fusion to overcome information deficiency caused by single face pose feature used in head pose estimation.The main idea is to extract the second-order gradient histogram feature of face image and the local binary model feature of uniform pattern from face image.Finally,the two extracted features are fused to construct a new fusion feature for head pose estimation.The new fusion feature is more sensitive in describing and distinguishing human face image against other single features.3.Head pose estimation for random forest improvement is set forth.This paper suggests an improved weighted-voting random forests algorithm so as to address the problem that equal-weight voting in random forests tends to spoil the overall performance of random forests.The adaptive algorithm was applied to each decision tree.As different decision trees have different functions,the weights of the decision tree in the final pose estimation voting should be updated to strengthen the classification accuracy of random forests.In this paper,head pose estimation is classified into two categories: coarse estimation and precise estimation according to different classification standards.The recognition rates of two are respectively 96.98% and 96.62%.The research results demonstrate that the head pose estimation has high recognition rate for the two types of estimation and achieve instant response.