一种对姿态稳健的鼻尖点快速定位算法

Pose-Invariant and Fast Method for Nose Tip Localization

提出了一种对姿态稳健的鼻尖点快速定位算法。在局部基准坐标(LRF)下计算顶点的平面距离能量,并设计了一种新的迭代筛选算法,计算得到候选点;计算候选点集中的每个顶点在人脸三维矢量场中的散度,将散度值最大的顶点作为鼻尖点。在FRGC v2.0和Bosphorus人脸库上对算法进行验证,在Bosphorus库上最终平均每张人脸定位仅耗时0.62s,在FRGC v2.0库上的定位准确率为95.6%。最后与当前其他算法进行对比,所提算法在速度和精度上均取得了较好的结果。实验结果证明所提算法不仅有望达到实时处理的要求,还具有较高的准确率,且对人脸姿态变化具有稳健性。

This study proposes a fast algorithm for nose tip localization, which is robust to pose variations. Based on the coordinates of the local reference frame (LRF), the plane-distance energy of each vertex is calculated and a novel iteration algorithm for selecting candidate points is designed. For each vertex with centralized candidate points, the divergence on the three-dimensional (3D) vector field is computed. The nose tip denotes the point with the maximum divergence value. The efficiency of the algorithm is verified by applying it to the FRGC v2.0 and Bosphorus face libraries. The average runtime of nose tip location is only 0.62 s on the Bosphorus library, whereas the location accuracy is 95.6% on the FRGC v2.0 library. Finally, compared with other state-of-the-art algorithms, the proposed algorithm ranks the first both in speed and accuracy. The results show that the proposed algorithm can meet the requirements of real-time processing, has relatively high accuracy, and is robust to the pose variations in human faces.