Research On Real-time Face Landmark Detection And Tracking Technology Based On Video

Face recognition technology is a kind of biometric recognition technology which extracts the human face appearance information to recognize the identity.It is a hot topic in the computer field and has been widely used in society and public life.Face landmark detection plays an important role in face analysis and recognition technology.It is also the basis and premise of facial expression recognition,face modeling and retrieval,facial identification and other related fields.Face landmark detection process is rather complicated,especially under the influence of uncertain external environment,such as skin color,pose change,illumination intensity and occlusion,the accuracy of traditional face landmark detection algorithm will decrease.At the same time,compared with static face landmark detection,face landmark detection and tracking in video environment is more challenging.Therefore,in the dynamic environment,determining the exact location of face landmark through detection or tracking technology is worthy further study.Therefore,the research work of this paper has the following aspects:1.Face landmark detection algorithm based on cascaded shape regression is studied in depth.On this basis,a double-level cascaded regression structure is designed to increase the stability of the regression algorithm.The gradient boosting regression tree algorithm is introduced,weak regressor is constructed by decision tree with simple structure.The residual of real value and predicted value is used as iteration fitting,which accelerates the iteration speed and effectively improves the accuracy of the algorithm.2.In view of the fact that the gradient boosting regression tree focuses on fast fitting of data,which may lead to over-fitting of prediction,the gradient boosting tree algorithm is replaced by the gaussian process regression algorithm.Because the common kernel functions do not match the decision tree algorithm,a special kernel function is designed to abstract the distance into whether the inputs falling into the same node,which reduces the complexity of training and prediction.According to the particularity of the kernel function,the algorithm of three-level cascade regression structure is designed.Using multiple decision trees and kernel function to regress the shape,the positioning accuracy is improved efficiently.3.In order to detect the landmark of human face in the dynamic environment of video,the detection-tracking method is adopted.The face is tracked by the algorithm based on kernel correlation filtering,and the multi-scale detection is added to solve the problem of fixed template.For the tracking drift problem in long-term tracking,a tracking verification system is designed,and local detection is used instead of global detection to reduce the system cost,so that the face feature point tracking system can achieve real-time and accurate detection.