Abstract: The original inverse compositional Active Appearance Model (AAM) only does one fitting process. When the initial position is far away from the destination, the model often falls into local minimum and becomes hard to converge into the correct position. Against this problem, a dual fitting method using Multi-Resolution AAM (MR-AAM) was proposed. Firstly, the first time fitting was to locate the initial position of the face in the low-resolution AAM, and the second time fitting was to use inverse compositional algorithm in the high-resolution AAM. This method can find the exact initial position and achieve better result of facial feature point localization. The experimental results show that the proposed method performs better than traditional method in the fitting accuracy along with the real-time case.

Key words: facial feature point localization, Multi-Resolution Active Appearance Model (MR-AAM), inverse compositional algorithm, dual-fitting, point-to-point error

摘要： 传统的主动表观模型(AAM)反向组合算法仅进行了单次拟合过程,当初始位置与目标对象偏移过大时,往往会陷入局部最小,难以收敛到正确位置。针对此问题,提出了一种基于多分辨率AAM(MR-AAM)的双重拟合方法,首先在低分辨率模型下进行第一次拟合以确定面部初始位置,然后在高分辨率模型下进行二次拟合。由于能够快速获得较准确的初始位置,进而取得较好的人脸特征标定结果。实验结果表明,所提方法与传统方法相比,在能保证实时的情况下,提高了拟合精度。

关键词: 人脸特征点定位, 多分辨率主动表观模型, 反向组合算法, 双重拟合, 点对点误差