非约束环境下人脸特征点的稳定跟踪

Stable Face Features Tracking under Unconstrained Condition

为了对人脸特征点进行精确地跟踪,提出一种在线参考表观模型(ORAM)的算法.首先在原主动表观模型(AAM)中加入在线更新的参考模型;然后采用子空间在线自更新机制,利用增量学习方法在线更新AAM的纹理模型和参考模型;在此基础上,基于同步反向合成建立ORAM的特征点拟合算法.为减少跟踪过程产生的累积误差,利用初始稳定跟踪结果建立纹理子空间重置机制,完成人脸特征点跟踪.实验结果表明,ORAM算法无需训练集,在姿态、表情、光照变化的环境下,能够准确、快速地完成人脸跟踪.

A method for tracking facial feature points stably is proposed to complete the facial feature points tracking accurately. First, the online update reference texture model is combined with the original active appearance model （AAM）. Second, use the subspace update mechanism. The AAM texture model and the reference model are updated via the incremental learning method. Then the online reference appearance model （ORAM） fitting algorithm based on simultaneously inverse compositional is designed. To reduce the cumulative error, texture subspace reset mechanism is introduced based on the first stably tracked frames. Finally, face features tracking is completed. Compared with other AAM algorithms that require a large amount of training data, ORAM need no training data. It is proved that this method can complete the face tracking accurately and quickly in different posture, facial expression and illumination conditions.