Improvement And Implementation Of Facial Landmarks Localization Algorithm Based On Constrained Local Model

In recent years, biological recognition technology is developing rapidly. As the most friendly recognition method of biological recognition technology, face recognition has drawn enormous attention and led to a lot of explorations in diverse research areas. In the method of face recognition, facial features with high recognition(such as eyes, nose, mouth and so on) are chosen as the key roles in facial landmarks localization, which is extremely important in full recognition. Precise facial feature points positioning will produce remarkable application effect in various areas like pose estimation, facial expression analysis, three dimensional face modeling and animation, etc. Based on these, the research of facial feature points localization has certain theoretical value and practical significance.With analysis of the existing algorithms and summary of commonly research methods in facial landmarks localization, this thesis mainly studies Constrained Local Models(CLM), analyzes the characteristics of shape model?local model of CLM and the corresponding modeling methods, fits optimizable parts after modeling, analyzes and improves classic fitting methods and Convex Q uadratic Fitting(CQF), indicates that there may exists local minimum problem, which may leads to fa ilure in finding the optimal location in local search, further influences global positioning accuracy. To solve this problem, the thesis introduces mean shift algorithm. Aiming at possibility of convergence failure led by fixed windows in mean shift algorithm, search window wide adaptive method is adopted. The improved mean shift algorithm is used in CLM to improve original algorithm.In this thesis, facial feature points localization system is improved based on OpenCV, original algorithm and improved algor ithm are used respectively in facial feature points under the natural environment(Label Facial Part in The Wild, LFPW). By comparing localization effect of the two algorithms in same situation, results show that improved algorithm solves the problems of local minimum, improves the feature point positioning accuracy, which has certain performance improvement.