Research On Face Detection And Tracking Algorithm Based On Video Sequence

Face detection and tracking is an important technology in the field of computer vision,scholars have done a lot of research on it.There are always some uncontrollable interference factors in the video sequence,such as illumination change,complex background,change of target scale and fast motion.Therefore,how to solve these problems and design some face detection and tracking algorithms with high accuracy and robustness is still a big problem.In this paper,three modules of face image illumination processing,face detection and tracking are respectively improved and optimized.The main research contents are as follows:Firstly,aiming at two problems in the face image decomposition process,such as the reflectance image is polluted by light,and the valuable discriminative information of the illumination image is lost,a new illumination processing method based on Local Edge-Preserving filter(LEPF)and illumination subspace is proposed.Firstly,the images are decomposed by LEPF to generate good reflectance images with invariant texture information.In addition,in order to effectively normalize the illumination image,a class-based illumination subspace is proposed.By projecting the illumination image onto the basic image in the illumination subspace,the variation of the illumination image can be extracted and eliminated effectively from the illumination image without losing its discriminating information.The experimental results show that the illumination normalization method proposed in this paper has some advantages compared with other algorithms under different illumination conditions.Secondly,aiming at some problems in the process of sample training,such as the training may be degenerated,and the weight of the positive and negative samples may be serious distortion,which may lead to reduce training accuracy and affect the detection rate.So Adaboost algorithm based on sample misclassification times is proposed,which is called MTAdaboost algorithm.It is found through experiments that the MTAdaboost algorithm can suppresses the over adaptation of the training target class weights.The combination of MTAdaboost and YCb Cr skin color segmentation is used in face detection,which can effectively improve the face detection rate and detection speed.Thirdly,in order to improve the accuracy of tracking algorithm in target scale or appearance change,a face tracking method based on feature point tracking is proposed.Because of the unbalanced distribution of corner points,the position of the tracking window deviates from the center of the face.So Cam Shift algorithm is adopted based on feature point tracking.Due to the face moving rapidly makes the face fuzzy,feature tracking method fails.So particle filter-based face tracking method is used.The initial position of Cam Shift is computed by particle filter,which improves the performance of capturing fast-moving face.In addition,a tracking mode detection method is proposed,which is used to determine what kind of tracking method is used in any case.By qualitative and quantitative analysis of the experimental results,the proposed method has a better performance under fast moving,target scale or appearance change.