基于多变量标签分布的连续型姿态估计方法

A Continuous Head Pose Estimation Method Based on Multivariate Label Distribution

人脸的头部姿态往往指示并传达着丰富的信息,准确估计头部姿态角度在人脸识别、表情识别等领域有重要作用。针对获得的人脸真实姿态角度往往存在一定的偏差且只包含有限个离散角度等问题,文中提出了一种基于多变量标签分布的连续型姿态估计方法。在训练阶段,对不同姿态角度,通过训练获得离散情况下的多变量标签分布;在测试阶段,采用正交多项式拟合的思想,将离散的分布拟合成连续的分布,计算分布的最大值所对应的标签作为最终的输出结果。文中在Pointing’04公开库上进行了测试,利用文中方法,在正交多项式拟合后,估计出的人脸的姿态角不再局限于训练集中的一些角度,而是有更多连续的值,所得的估计姿态角更接近于真实角度。实验结果表明,文中方法能够预测出更多的人脸姿态角度,并且预测更稳定。

The human＇s head pose are abundant of information. Accurate estimation of head pose plays an important role in face recognition,expression recognition and so on. To improve the precision of estimation and to alleviate the problem that poses are always fixed to some angles,a continuous method based on multivariate label distribution to estimate head poses was presented. In the training phase,get the discrete multivariate distribution from discrete poses and angles. In the testing phase,adopt orthogonal polynomial fitting to transform the discrete distribution into continuous distribution and compute the label corresponding to maximum in distribution as final output. The proposed method has been tested on the open Pointing ＇04 database. After orthogonal polynomial fitting,the estimated angles are no longer limited to angles in test set,but more continuous values. The estimated angle obtained by the method proposed is closer to the real angle. The result indicates that this method can estimate head pose in wider angle,the result is more stable.