基于特征点表情变化的3维人脸识别

目的 为克服表情变化对3维人脸识别的影响，提出一种基于特征点提取局部区域特征的3维人脸识别方法。方法 首先，在深度图上应用2维图像的ASM(active shape model)算法粗略定位出人脸特征点，再根据Shape index特征在人脸点云上精确定位出特征点。其次，提取以鼻中为中心的一系列等测地轮廓线来表征人脸形状；然后，提取具有姿态不变性的Procrustean向量特征(距离和角度)作为识别特征；最后，对各条等测地轮廓线特征的分类结果进行了比较，并对分类结果进行决策级融合。结果 在FRGC V2.0人脸数据库分别进行特征点定位实验和识别实验，平均定位误差小于2.36 mm，Rank-1识别率为98.35%。结论 基于特征点的3维人脸识别方法，通过特征点在人脸近似刚性区域提取特征，有效避免了受表情影响较大的嘴部区域。实验证明该方法具有较高的识别精度，同时对姿态、表情变化具有一定的鲁棒性。

Expression-insensitive 3D face recognition method based on facial fiducial points

Objective Compared with other biometric recognition methods, face recognition is friendlier, more natural, and less interferential to users. Hence, face recognition has numerous applications in various fields. However, the 3D shape of the human face is a nonrigid free-form surface, which causes face region distortion under expression variations, particularly in the mouth region. Therefore, 3D face recognition is easily affected by facial expression variations. Laughing and acting surprised make the face produce a hole in the mouth region and consequently change the topology of the facial model. This study proposes a 3D face recognition method based on facial fiducial points to avoid the effect of different facial expressions. Method The active shape model algorithm, which is usually used for 2D images, is applied to roughly detect facial fiducial points in depth images. The shape index is used to accurately locate fiducial points in 3D point clouds. A series of iso-geodesic contours from the landmark (located in the middle of nose tip and nose root) is then extracted to represent the facial shape and avoid the mouth region. Procrustean features (distances and angles) defined by pose-invariant curves are selected as the final recognition features. The classification results of each geodesic contour are compared and combined at the decision level as the final results. Result FRGC V2.0 is a large public face dataset. Experiments on the detection and recognition of facial fiducial points are conducted using the FRGC V2.0 dataset. In the detection experiment, face models in fall 2003 are selected as the training set, and 150 scans acquired in spring 2004 are selected as the testing set. Seven fiducial points, including eye corners, nose tip, and mouth corners, are manually located in the testing set. The accuracy of the seven detected fiducial points is measured through the Euclidean distance between the manually detected fiducial points and the corresponding automatic points. Each of the seven points is accurately detected, and the mean value of the positional error is less than 2.36 mm. In the recognition experiment, 424 scans from 60 subjects in the FRGC V2.0 database are randomly selected. Each iso-geodesic contour with procrustean features (distances and angles) is tested, and the classification results of the eight iso-geodesic contours are weighted in a decision-level fusion. The final Rank-1 recognition rate is 98.35%. Conclusion Expression variation is an important research direction for 3D face recognition. This study proposes a novel method for locating fiducial points in 3D point clouds. Combined with the depth images and 3D point clouds, seven fiducial points are completely, automatically, rapidly, and accurately detected. A series of facial contours based on these points is extracted to represent the facial surface. The effect of expression variations on recognition is decreased because the iso-geodesic contours are located in approximate rigid regions. Overall, the proposed method is valid and robust in the presence of pose and expression variations.