Facial Feature Points Tracking with Online Reference Appearance Model

To complete facial feature points tracking accurately, we propose an improved active appearance model (AAM), which is online reference appearance model (ORAM) with the subspace update mechanism. The online update reference texture models are combined with the conventional AAM. When the model fitting is satisfactory while tracking, we update the AAM texture model via the incremental learning method and simultaneously update the reference texture model with the current frame appearance information. The ORAM-fitting algorithm is established based on simultaneously inverse compositional (SIC). To reduce the cumulative error produced in the tracking process, we provide the texture subspace reset mechanism with the original stability frames. Compared with other AAM algorithms that require multiple training images, ORAM achieves accurate tracking results with no training data needed. Experiments indicate the proposed method can complete face tracking accurately and quickly.