自底向上的人脸特征点定位

Human face detection and feature tracking in a bottom-up way

针对人脸特征点定位的复杂性,提出了一种鲁棒的人脸特征点定位方法.该方法通过对497张不同年龄层次、不同性别、不同脸型的人脸进行训练学习,可以自动地对新输入的人脸图像进行识别和定位.与传统方法不同的是,该方法首先通过皮肤模型确定人脸的大致位置,然后采用自底向上的方法,利用主动外观模型(AAM)对特征明显的局部器官进行定位,再进一步定位器官组合,直至完成整张脸的器官定位.与传统方法相比,该方法不是局限于全局化的定位手段,而是通过逐层调整匹配精度的形式,使总体匹配精度得到提高.

A robust automatic tracking method of human face feature points was proposed to enhance the facial feature tracking performance. By training the algorithm with 497 different faces of different ages, genders and face shapes, the algorithm first computed a coarse position of the human face by a trained skin model. A bottom-up scheme was applied to find the face components with obvious features that were firstly tracked by active appearance model （AAM）. These components were organized as combinations of parts and then all of face organs with different weights to enhance the recognition and tracking performance. Compared with traditional methods, this method no longer uses a unified locating method, but a layer-bylayer one, and adjusts the accuracy in each layer, thus generates better tracking results.