Research On 3D Object Recognition Based On Local Feature

3D object recognition is a very important basic research in computer vision.It is the premise and basis for the research of remote sensing,biological medicine,robotics,and many other fields.So 3D object recognition has a wide application prospect.Existing 3D object recognition methods can be divided into two categories,global feature based methods and local feature based methods.The global feature based methods ignore the shape details and they are therefor not suitable for the recognition of partially visible object from cluttered scenes.And the technology of 3D object recognition based on local feature has more advantages.This kind of method comprise three phases: 3D feature points detection,feature points description and the local surfaces matching.This paper take the three phases as the study's core and has detailed in-depth study to 3D object recognition based on local feature.The main work is summarized as follows:(1)Aiming at the problem of scale invariant feature and traditional algorithm to noise sensitivity,a multi-scale feature extraction algorithm based on scattered point cloud is proposed.Multi-scale analysis of scale space is constructed by changing the local neighborhood size.In the different scales,the variation of the surface is calculated by the covariance analysis of the local neighborhood,and the feature points with scale invariance are found.At the same time,the point signature method is introduced to enhance the robustness of the noise.(2)For traditional descriptor,the matching time is longer and the dimension is so large.In this paper,a geometric covariance descriptor is proposed,and the covariance matrix is constructed to describe the feature points by using the geometric features such as the angle and distance between the feature points and the neighborhood points.The experimental results show that the descriptor not only has strong description ability,but also has the property of rotation translation invariance.and is not sensitive to noise and variation of sampling density.(3)In surface matching process,aiming at the problem that there is a certain mismatch between the nearest neighbor matching of feature points,this paper introduces the canonicalcorrelation analysis to eliminate the false matching of feature points.Finally,a good matching effect is obtained.