基于全脸统计学模型和回归器的对遮挡鲁棒的三维人脸特征点定位方法研究

项目名称： 基于全脸统计学模型和回归器的对遮挡鲁棒的三维人脸特征点定位方法研究

项目编号： No.61303121

项目类型： 青年科学基金项目

立项/批准年度： 2014

项目学科： 自动化技术、计算机技术

项目作者： 赵玺

作者单位： 西安交通大学

项目金额： 23万元

中文摘要： 特征点定位，是三维人脸分析研究工作的基础性问题之一。现有的特征点定位方法在准确性、鲁棒性及计算效率上都有待提高；尤其是针对带有遮挡的三维人脸，目前缺乏准确有效的定位方法。 本项目将研究带遮挡的三维人脸特征点定位方法，其基本思路：基于一个无遮挡可形变全脸统计学模型和局部三值模式，对遮挡进行精确的检测和遮挡物体移除；采用改进的拟合算法将可形变全脸统计学模型拟合到带有缺失的人脸上，从而对移除了遮挡的人脸部分进行数据修复；研究并实现基于级联回归器的定位算法，并对不同的人脸特征训练和回归器假设进行对比分析，以达到对三维人脸特征点定位的最佳结果。 本项目的目标：在带遮挡的条件下，使三维人脸特征点定位的精确度、鲁棒性和实时性得到显著提高。

中文关键词： 三维人脸；特征点定位；遮挡；三维重建；表情识别

英文摘要： The automatic three-dimensional facial landmark localization problem is one of the fundamental problems in 3D face analysis. However, the current 3D facial landmarking methods need improvement in aspects of accuracy, robustness and efficiency, especially to those data in the presence of occlusions. This project targets on the problem of localizing the landmarks on 3D faces with occlusion. It attempts to detect and remove the occlusions on the facial regions using a holistic deformable statistical facial model with aid of Local Ternary Patterns; attempts to fit this deformable facial model to the 3D facial data with missing regions after the occlusion removal, so that these empty regions can be impainted and recovered; attempts to study and develop a landmark localization algorithm based on a cascaded regression model; and attempts to systematically evaluate and compare the landmarking algorithm's effectiveness on different facial feature banks in both texture and shape representations and different regression hypotheses, so that the best localization performance can be achieved. The objective of this project is to develop a new automatic facial landmark localization method with improved accuracy, robustness and efficiency compared to the state-of-the-arts landmarking techniques, especially for those facial data

英文关键词： 3D face；landmark localization；occlusions；3D face reconstruction；facial expression recognition