综合鲁棒特征和在线学习的自适应三维人脸多特征跟踪

Adaptive 3D Facial Feature Tracking Combining Robust Feature with Online Learning

提出一种灰度与边强度信息相结合的鲁棒特征并综合在线学习方法来进行自适应视频人脸多特征跟踪。算法思想是利用三维参数化网格模型对人脸及表情进行建模,利用弱透视模型对头部姿态建模,求取归一化后的形状无关灰度和边强度纹理组合成一种鲁棒特征,建立单高斯自适应纹理模型,并采用梯度下降迭代算法进行模型匹配得到姿态和表情参数。实验证明,本方法比单纯利用灰度特征在复杂光线和表情下具有更好的鲁棒性。

An algorithm based on robust feature combining edge strength and raw intensity and online appearance model fitting was proposed to track head pose and facial actions in video. A 3D parameterized model, CANDIDE model, was used to model the face and facial expression, a weak perspective projection method was used to model the head pose, an adaptive appearance model was built on shape free intensity and edge texture, and then a gradient decent model fitting algorithm was taken to track parameters of head pose and facial actions. Experiments demonstrate that the algorithm is more robust than using only intensity especially when the lighting condition and facial expression is complicated.