Research On Human Ear Detection And Recognition Based On 3D Point Cloud Model

Ear biometrics is proved to be an effective method used in identification verification in recent years. Compared with existing biometric method, human ear, the research object in ear biometrics, has stable structure and abundant spatial details for feature extraction. Besides it can be sampled in no-invasive ways, which is preferable in practical use. In order to hold ear's rounded spatial structures, we introduce 3D reconstruction method to image human ear. Work presented in this paper consists of designing a 3D point cloud acquisition system for ear imaging based on laser scanning, 3D point cloud model preprocessing, feature extraction and further research into 3D recognition.In order to reconstruct ear model, 3D range scanner with extraordinary high price is prerequisite to existing research. Subsequently ear detection and recognition is made possible to go ahead with based on 3D ear imaging. For the purpose of reducing expenses, we design a set of 3D imaging device and corresponding capture software with relatively low cost which still do not decrease recognition precision. Constituted by components that are readily available and inexpensive, this set of device is based on strip laser scanning according to principle of structure light. Each pattern of laser lines on the trace of scanning orientation is recorded for 3D reconstruction and feature extraction, which are prerequisites for recognition.3D preprocessing to the point cloud of ear consists of noise reduction, ear detection and posture normalization. As posture of ear changes to the motion of head, we firstly eliminate isolated noise as well as popcorn noise, then apply a new method to segment ear region of interest from surrounded facial skin through 3D curvature estimation and finally propose a neoteric 3D coordinate direction coordination method focusing on ear ROI to unify ear's posture. In this schema, projection pursuit helps to adjust point cloud's projecting direction and ensures the sparsest point distribution on the projection plane, while principle component analysis unifies rotation angle on projecting direction if it has been worked out. This method leads to faster registration and provides more convenience for feature extraction.Point's spatial and curvature distributions are both appropriable features for recognition. We introduce rough point set and point cloud's slice projection for research into recognition. Traditional 3D point registration method: iterative closest point method is improved in our experiment with assist of a database including a number of individual samples captured in previous work. Finally the recognition results are evaluated and we make comparison between different methods.