Research On Face Tracking Feature Point Position

Currently the computer vision favored by many scholars. Video based face tracking and recognition become a research hotspot in the field of computer vision. It has a broad application prospect. In the security monitoring system, video conference, entrance guard system, remote teaching, and identity recognition, it will be all used. Face detection, tracking, and identification are inseparable complements each other. Face detection is the basis for face tracking, and provide information to track the target initial position. Face detection and tracking of feature points provide accurate location information for the subsequent fitting. This paper comprehensively researches three parts:the face detection, Video based face tracking and facial feature points localization.In the part of face detection the Adaboost-algorithms will be applied.First of all, it is needed to train face classifiers. There are five kinds of Haar-like characteristics to be used in showing the level of face gray distribution. Integral figure method is adopted to simplify the calculation. In the training process, the renewal of the sample is realized by changing the weights of samples. Training after T cycles, classifiers with the weight, of which the Classification is relatively weak, are produced.Then the classifier, of which the Classification is relatively strong is produced by the weighted combination of several weak classifiers. After that, several classifiers, of which the classification capability is progressively enhanced, are cascaded to construct a cascade-classifier. After completing the above steps, face-detector training is completed. In order to improve the accuracy and robustness of face detection, in this paper, some measures are adopted,such as illumination compensation、histogram equalization and noise filtering operation. After the preprocessing, Face detection in the entered picture can be executed by using Adaboost face-detector. It is in preparation for the subsequent face-tracking and recognition.In this paper, the Video based face tracking is carried out based on Adaboost Algorithm of face detection. Traditional Camshift algorithm is a kind of semi automatic human face tracking, and We need to manually enter the initial position of the face at the beginning of the track.In this paper, through combining the Camshift algorithm with the Adaboost algorithm, It has realized the automation of the face tracker. Face tracking is based on human face skin color, vulnerable to light、the interference of background and the target similar to skin color, and these disturbances will affect the tracking effect. In response to those problems, this paper gives the corresponding improving solutions. Firstly, the Adaboost detector is applied. Before tracking, the detector is enabled to locate human faces and then given an initial position to search window. Duing the tracking process, if tracking target is missing caused by the too large distance between two frames、some obstructions or the target similar to skin color appearing, the tracking system can re-enable Adaboost face tracker for repositioning. Through the introduction of Adaboost face detection device, We achieve a good efficiency of Camshift algorithm and get a good solution to the problem of tracked targets missing, and automatic detection and tracking performance cycle of the tracking System is realized. Secondly, in the probability density distribution diagram, there are black pixels within the face region. In order to make the image more connected and remove noise points of the face, In this paper, it put the Median filtering function cvSmooth to use. It also adopts corrosion and expansion method to improve the original algorithm. Additionally, as traditional Adaboost detection algorithm uses the rectangle to mark the detected face position, not meeting the face shape, some background points In the corner will be introduced in. In this paper,we convert it to an oval and add a rotation angle to it, Effectively improving the accuracy and robustness of the tracking algorithms. Face tracking position provides the initial fitting parameters for the subsequent feature fitting.In this paper, Active Appearance Model (Active Appearance Model, AAM) in combination with a reverse facial feature is put use to facial features fitting. Firstly, the initial position parameters of the feature fitting are initialized by the positioning result of the Meanshift algorithm. Then the shape model is driven continuously close to the position of the input picture face by the energy function of the texture to the true until the energy function reaches a minimum. At this point we need to solve out local and global shape parameters and texture parameters in order to locate the key feature points. The traditional reverse combination AAM fitting algorithm can achieve good results in the application of frontal face images positioning. But for the positioning of the faces that have a rotation angle in the plane, the accuracy and efficiency are not very high. In order to solve that problem, this paper proposes to strike a location in the face of the eyes to get the rotation angle of the face. Then te angle of rotation is assigned to the initial position parameters of the model. This solution not only reduces the number of iterations, but also improves the accuracy of the key features location. In some of the original pictures fitting failure, this algorithm can accurately find the location of key feature points.In this paper, the experimental platform of Visual C++6.0 embedded with Opencv1.0 is used to verify the correctness of theoretical analysis.