Research On Point Cloud Matching Algorithm For Multiple Posture Hand Vein Recognition

Compared to biometrical recognition method based on 2D vein image,using 3D point cloud to express hand vein structure can ensure the absolute scale of the vein.3D Vein point cloud is not affected by the acquisition device and the change of hand posture.Moreover,3D vein point cloud has the depth information,which can provide a greater discrimination in recognition.However,the existing vein point cloud recognition algorithms are essentially based on point-to-point matching,which are sensitive to hand posture and time consuming.Based on the analyses of the characteristics of hand vein point cloud,the thesis focuses on the point-to-model vein point cloud registration algorithm to achieve multi-pose hand vein recognition.A point cloud registration algorithm based on Gaussian mixture model is proposed to match hand vein point cloud,which combine with the initial registration method based on 3D feature array,and the normal distribution transform algorithm is used to evaluate the spatial coincidence of two point clouds to complete the authentication as well.The major works of this thesis is as follows:1.Point-to-model point cloud registration algorithm is introduced into vein point cloud matching.The feasibility of the normal distribution transform algorithm in vein point cloud matching is verified.Based on this,a point cloud registration algorithm based on Gaussian mixture model is proposed,which can effectively solve the effect of hand posture changes in hand vein recognition.The normal distribution transform algorithm is used to evaluate the coincidence of two point clouds and to realize the multi-pose hand vein recognition.2.An initial registration method based on 3d feature array is proposed for coarse alignment.In the stereo vision system,the SIFT features are reconstructed into three-dimensional features for initial point cloud registration,which avoids the possibility of Gaussian mixture model-based algorithm falling into a local extreme when the difference of the posture between two point cloud is too large.3.CUDA universal parallel computing architecture is used to design and implement parallel computing improvements based on GPU,which achieved great acceleration compared with CPU version.4.A hand vein point cloud acquiring system based on stereo vision is constructed and the 3D vein point cloud is reconstruction according to the stereo-vision principle.A vein point cloud database is established,in which 50 experimental hands vein with 13 different postures of each hand were collected.To verify the proposed algorithm in multi-pose hand vein recognition,the veins cloud authentication experiments are designed and conducted based on the 650 point clouds data.Cross matching experimental results show that when the hand posture changes limited to the range ± 20 degrees,the equal error rate of the system is only 1.69%.The theoretical analysis and experimental results show that the proposed method has the advantages of fast authentication and low initial pose requirements for point cloud.The method can effectively overcome the influence of hand pose changes on the recognition results to some degree,which indicate an inspiring and promising idea in multi-pose hand vein recognition.The recognition rate is larger than 98% even when the posture of the hand changes within a range of 40 degrees while the average certification time is less than 500 ms for 1:1 authentication.