Research On Local Descriptors In 3D Objects

The popularity and price reduction of 3D image acquisition devices(such as Microsoft's Kinect and Intel's RealSence)make the 3D data of objects more accessible,which lays a good foundation for the study of all aspects of 3D objects.Research related to 3D objects has gradually become a hot spot and has important applications in robot vision,autopilot,3D face recognition,3D object positioning and 3D modeling.Feature extraction is a very important stage in the research of 3D objects.3D objects features are generally divided into global features and local features.Global features refer to the features extracted from 3D objects as a whole,but need to be advanced.The three-dimensional object is segmented and the features extracted under various scenes are unstable.The feature points are extracted first on the three-dimensional object,and then the features extracted on the feature points are called local features.The local features are more robust in the case of occlusion.This paper is based on local features.The local descriptor is the abbreviation of local feature extraction method.The feature extraction is generally based on the spatial distribution information and geometric information of the nodes around the feature points,but there is not yet a good recognition effect,simple extraction method and In this case,the local descriptors are stable.In addition,in some applications,some machines also have memory limitations and speed requirements,while the general floating-point descriptor memory usage is relatively large.Based on this,there are two main contents of this paper.Firstly,the TOLDI is improved.the local reference frame use more neighbor nodes to construct LRF,and neighbor nodes are weighted according to the distance between the neighbor nodes and the feature points.The improved local coordinate system is more repeatable under normal scenes and can achieve better results under some occlusion situations.Secondly,the angle information is added,on the one hand,it can prevent the error matching situation in the matching process,on the other hand,the feature is enriched and the calculation is not too complicated.Secondly,a binary descriptor is proposed to solve the problem of time efficiency and memory usage.Firstly,the local coordinate system establishment method of Signatures of Histogr ams of Orientations(SHOT)is improved,so that the improved local coordinate system is faster and more repeatable.Secondly,based on the triple orthogonal local depth descriptor,the spatial distribution information of the points is added at the same time,and then the obtained feature matrix is re-extracted with the binary descriptor,and the resulting descriptor takes up more space than before.A lot smaller.At the same time,we use the binary descriptor matching method(Hamming distance)to match,which greatly reduces the time used for matching.