基于人眼优先拟合的AAM人脸特征点跟踪

Facial feature tracking of AAM algorithm based on eyes first fitting

使用主动外观模型(active appearance model,AAM)对人脸特征点进行跟踪时,当人脸姿态出现大幅度偏转以及初始位置与目标人脸偏离较大时,将导致人脸特征点跟踪失败。针对以上问题,应用支持向量机算法估算出当前人脸的偏转角度并对人脸姿态偏转模型中的参数进行实时更新,有效解决人脸姿态大幅度偏转问题。使用强跟踪卡尔曼滤波算法进行人眼跟踪,将所获取的当前人眼位置坐标与人脸姿态相结合优先对人眼进行特征点拟合,当人眼特征点拟合完成后再对剩余人脸部件特征点进行拟合,提高了人脸特征点跟踪的稳定性与实时性。最后通过实验表明算法在视频人脸特征点跟踪的准确性、实时性和鲁棒性方面具有良好的性能。

The application of AAM would cause the tracking failure of facial feature point when face pose greatly deflected and initial position deviated drastically from the target face. To solve the above problem, this paper applied SVM to estimate the deflection angle of the present face and update the parameters of deflection model of face pose in real time, thus efficiently settled the matter of great deflection of face pose. This paper used strong tracking Kalman filter algorithm to track eyes, and combined the present eyes position coordinates with face pose so as to fit the eyes feature points firstly. After the fitting of the eyes feature points, the left facial components feature points would be fitted. This algorithm improved the stability and efficiency of facial feature point tracking. The experimental result shows that the proposed algorithm has good performance in the accuracy, efficiency and robustness of facial feature point tracking in videos.