Research On Part Point Cloud Construction And Multi-view Registration Based On Binocular Structured Light

The 3D point cloud of part is a very important data source in the process of industrial detection and measurement.The general construction method of 3D point cloud based on contact measurement has high cost and complicated operation.Although binocular stereo vision technology can build point cloud data at a lower cost,but due to the limitation of pixel stereo matching,the construction effect is poor for the area lacking texture inside the part.The generated point cloud is sparse and low-precision.In recent years,with the development of 3D sensors,especially the progress of binocular vision and structured light coding technology,it is possible to construct high-density,high-precision and high-efficiency point cloud of part.However,there are still three problems to be solved in the point cloud construction and registration based on binocular structured light:(1)In the aspect of part point cloud construction,due to the shielding of binocular field of view,the traditional stereo matching algorithm has the phenomenon of mismatching at the edge of the part in the image,which affects the effect and quality of the construction of the part point cloud.(2)In the aspect of point cloud preprocessing,environmental factors such as part fixtures and limiting devices in the scanning field form high-density redundant clusters of point cloud.It is necessary to design the corresponding de-redundancy algorithm to eliminate it.(3)In the aspect of multi-view registration of part point cloud,how to establish the mapping relationship between the turntable angle and the rigid transformation of point cloud with high accuracy and flexibility is the current research difficulty and hot spot.In order to solve the above problems,based on binocular basic model and structured light coding technology,this paper conducts an in-depth study on the construction of part point cloud and the method of multi-view registration.The specific contents are as follows:(1)In the aspect of part point cloud construction,this paper presents an overall algorithm flow for the construction of part point cloud.Firstly,the image distortion correction and pole correction are carried out.Then the effective region of the part in the image is extracted.Secondly,absolute phase value is solved according to a series of structured light image.In the process of binocular stereo matching based on absolute phase,aiming at the pixel mismatching phenomenon caused by the occlusion of visual field at the contour edge of the part in the image,an absolute phase stereo matching algorithm based on threshold screening is proposed in this paper.Through reasonable threshold construction strategy,the mismatched pixels in the contour edge region are filtered out.Finally,parallax subpixel interpolation is used to improve the effect and quality of point cloud construction.(2)In the aspect of part point cloud preprocessing,the outlier noise data is filtered based on statistical filtering.Factors such as fixture and limit caused high-density point clouds block,this paper presents a method to remove high-density point cloud block based on Euclidean clustering segmentation,Firstly,the data structure of point cloud index is established based on KD-Tree.Secondly,the clustering segmentation of point cloud is realized based on the Euclidean distance between points.Then,a reasonable scale factor is designed based on the number of classification points to eliminate redundant point cloud block data.Finally,for the burr noise mixed inside the point cloud of the part,this paper adopts the method of moving least squares to smooth the point cloud and improve the surface quality of the point cloud of the part.(3)In the aspect of part point cloud registration,a multi-view registration method of part point cloud based on the reference angle of turntable is proposed.Firstly,the point cloud of calibration corners is registered before and after rotation.Then the rigid transformation matrix of the point cloud corresponding to the reference angle is calculated.Secondly,the rigid transformation mapping model between the turntable angle and point cloud is constructed through the recursive relation to realize the rough registration of the multi-view point cloud.Finally,a global incremental precision registration strategy was designed based on ICP algorithm,which modifies the coarse registration point cloud and realizes the precise registration of the part point cloud.Multiple repeated experiments are carried out with the standard measuring ball and parts of step shaft and step table.The experiment results show that the accuracy of the single-view point cloud of the part constructed in this paper can reach within 0.078mm,and the accuracy of the multi-view point cloud registration can reach within 0.058mm.Therefore,the point cloud construction and registration algorithm proposed in this paper can accurately and effectively reconstruct the three-dimensional information of the part.