基于spin image的人脸点云特征定位

Characteristics positioning of facial point cloud based on spin image

展示一个三维人脸点云特征点定位方法。通过HK(平均曲率和高斯曲率)这一曲面形状描述方法,划分人脸特征点候选区域,对特征点形成的旋转图像(spin image)进行比较,实现任意姿态下的鼻尖点和左右内眼角点的定位。原始的旋转图像算法计算量太大,运算时间过长,无法获得实际应用,在原算法的基础上,只对筛选出来的少量重要特征点进行区域曲面重构,避免大量无意义的点运算,提高旋转图像算法的实时性能。在GavabDB数据库上的实验结果表明,该方法最高获得了95.37%的识别率,对姿态、表情变化具有一定的鲁棒性。

A method to locate 3D facial features based on point cloud was outlined using spin images. The candidate regions of facial features points were divided using HK （mean curvature and Gaussian curvature）, which was a surface shape description method. The spin image formed by characteristic points was compared to positioning both inner corner of eye point and the tip of nose in any attitude. The initial spin image algorithm was hard to be used in real applications because of the large amount of computation and long calculation time, a method based on the initial spin image algorithm was proposed, in which only a small number of selected important features points were calculated and compared to reconstruct the regional curved surface, thus a large number of meaningless point operations were avoided and the real-time performance of the spin image algorithm was improved. Experimental results on the GavabDB database show that the method obtains a recognition rate of 95.37%, and it has certain robustness against changes of postures and facial expressions.