Research On Fast 3D Object Recognition Method Based On Point Pair Features

Object recognition from 3D point cloud is the core task in the field of computer vision and pattern recognition.It is also the key to the automatic operation of industrial robots in recent years,so it has attracted more and more researcher’s attention and become a new research hotspot.This dissertation makes a comprehensive analysis of the current 3D object recognition algorithms,aiming at the problems of the 3D object recognition algorithm in recognition speed and recognition rate under occlusion and clutter.In-depth study on 3D point cloud object recognition algorithm is made.Uniform downsampling algorithm of 3D point cloud based on unequal number of voxel grids,object placement plane extraction algorithm based on point pair features and convex hull solution,adaptive uniform point selection for 3D object recognition and distribution uniformity evaluation method based on Hausdorff distance,fast 3D object recognition method based on plane-constrained point pair features are proposed.The fast recognition of 3D objects placed on a plane under the condition of occlusion and clutter is realized.The main contents of this dissertation are summarized as follows.1.This dissertation investigates the theories and algorithms of 3D object recognition in recent years,summarizes the research status and progress of 3D object recognition at home and abroad,and determines the key problems,research contents,and research plan of this dissertation.2.To solve the problem of uneven distribution of sampling points in Drost’s voxel grid downsampling method,a uniform downsampling method based on unequal number of voxel grids is proposed.In this method,the edges of the axis-aligned bounding box along the x,y and z axes are divided into unequal number of parts,so that each voxel of the bounding box is approximately a cube,instead of a cuboid similar to the bounding box in Drost’s method.The experimental results show that the point cloud distribution obtained by this method is more uniform than that obtained by the Drost’s voxel grid method,and the computational efficiency of the proposed method is higher than that of the Drost’s voxel grid method.3.Aiming at the problems of slow extraction speed and low extraction accuracy of existing plane extraction algorithms,a plane extraction algorithm based on point pair features is proposed.The proposed algorithm calculates the point pair feature descriptors between any two points,and uses the properties of point pair features in the same plane to judge whether any two points are coplanar or not.On this basis,an object placement plane extraction algorithm based on convex hull solution is proposed to extract the object placement plane from the scene point cloud.The experimental results show that the object placement plane extraction algorithm based on point pair features and the convex hull solution can efficiently and accurately extract planes in the scene cloud,and accurately extract the object placement plane on this basis.4.In the scene feature description phase of the point pair feature 3D object recognition algorithm,the common sampling methods have the problem that the sampling parameters need multiple trial and error before final determination and the selected points are unevenly distributed.To solve this problem,an adaptive uniform point selection method is proposed.The proposed method can select the appropriate number of traversal points in the scene adaptively according to the size of the target object to be recognized,so as to ensure that there are traversal points falling on the target objects to be recognized in the scene.Aiming at the problem that the existing point cloud uniformity evaluation method is not suitable for the distribution uniformity of the selected point cloud in the original point cloud,a point cloud distribution uniformity evaluation method based on Hausdorff distance is proposed.The method transforms the distribution uniformity of the selected point cloud in the original point cloud into the similarity problem of measuring a point set and its subset.The similarity problem can be realized by calculating Hausdorff distances between the point set and its subset,thus solving the problem of evaluating the distribution uniformity of the selected point cloud in the original point cloud.The experimental results show that the distribution uniformity of a point cloud selected by the proposed method is better than the existing systematic and random sampling methods.Moreover,the point cloud distribution uniformity evaluation method based on Hausdorff distance can better evaluate the distribution uniformity of a point cloud.5.Aiming at the problem that the recognition speed of the point pair feature recognition algorithm needs to be improved,a fast plane-constrained point pair feature 3D object recognition algorithm is proposed.By eliminating all the points in the object placement plane and under the object placement plane in the scene point cloud,the number of point pair feature descriptors to be calculated is greatly reduced.By adaptively and uniformly selecting points,it is ensured that there are traversal points on the target objects in the scene point cloud,so that the prerequisite for the target objects in the scene to be recognized can be met.Then the point pair feature descriptors between traversal points and all other points in the scene are calculated,and then feature matching,pose clustering and ICP refinement are carried out to obtain the pose transformation matrix of the object to be recognized.The experimental results show that the algorithm based on plane-constraint point pair features can significantly improve the recognition speed.6.On the basis of the above theories and algorithms,the algorithms such as uniform downsampling,object placement plane extraction and adaptive uniform point selection are implemented on OpenCV and PCL platforms.Taking two scene point clouds with and without occlusion as examples respectively,the recognition effect of Drost’s point pair feature algorithm is compared with that of the proposed algorithm.Experimental results show that,compared with Drost’s algorithm,the proposed method can recognize and localize the target object quickly and correctly.Furthermore,when the object is occluded and there is clutter,the proposed method can still recognize and localize the object correctly.