3D face recognition based on a modified iterative closest point method

Face recognition has gained extensive attention recently, with many applications in a broad range of domains such as access control in security systems and picture tagging in social network web sites. This project builds a 3D face database and recognizes the unknown 3D face images in comparison with the 3D face database. In 3D face images used in this thesis are acquired by a 3D data acquisition system based on Digital Fringe Projection Profilometry (DFPP). DFPP is an efficient 3D data acquisition system to capture 3D data, with its simple system structure, high resolution and low cost. The 3D database consists of thirty group images In each group, there are three images corresponding with three views with (i.e. left-side view, right-side view, and frontal view) at the same scale of the same subject. The scale is different from group to group. To achieve 3D face recognition, there are two parts devised: image alignment and comparison. In order to implement efficient and accurate image alignment, two steps which are coarse alignment and fine alignment are implemented. In the coarse alignment step, two 3D images are roughly aligned into a same coordinates system and roughly aligned. After the coarse alignment step, the two face images will be aligned closer and an initial estimated value will be given for the fine alignment. A modified partial Iterative Closest Point (ICP) method is proposed in the fine alignment step. The partial ICP method is an efficient alignment method for 3D data reconstruction and 3D face recognition. It iteratively aligns the two point sets based on repetitive calculation of the closest points as the corresponding points in each iteration. However, if two 3D face images with different scales are from the same person, the partial ICP method does not work. In this thesis, the scaling effect problem of 3D face recognition has been solved. A 3×3 diagonal matrix as the scale matrix in each iteration of the partial ICP has been well designed. The probing face image which is multiplied by the scale matrix will keep the similar scale with the reference face image. Therefore even if the scales of the probing image and the reference image are different, the corresponding points can be accurately determined. The mean square distance between the two face images are compared to recognize that whether the two face images are from the same person or not. Based on the experiment results, the 3D face recognition can be achieved via the method proposed in this thesis. The mean square distance between two face images from the same person can reach to less than 0.05 while the two face images from the different persons can only keep 0.10 to 0.30