Complex special-shaped curved surface three-dimensional segmentation method and system based on robot vision

Abstract

The invention discloses a complex special-shaped curved surface three-dimensional segmentation method and system based on robot vision. The method comprises the following steps: after acquiring pointcloud data of a special-shaped curved surface workpiece through a three-dimensional laser camera, firstly, performing point cloud preprocessing in a working range on point cloud data based on a straight-through filtering method; eliminating a large number of irrelevant background point clouds; then, adopting a voxel grid down-sampling method to carry out sparse processing on the remaining point clouds, so that the point cloud segmentation rate is further increased; and finally, based on a KPConv method, carrying out example segmentation on the curved surface point cloud containing the tail endof the manipulator. On the basis of a Kernel Points point cloud convolution kernel, a new convolution method locally depending on a geometrical characteristic relationship between points is designed.The local point feature extraction capability is enhanced, the deformable convolution kernel of KPConv and the local point cloud feature extraction capability are integrated, and the method and the system have the advantages of being high in efficiency and accuracy and good in segmentation effect.

Description

technical field [0001] The invention belongs to the field of three-dimensional point cloud segmentation, and in particular relates to a method and system for three-dimensional segmentation of complex special-shaped curved surfaces based on robot vision. Background technique [0002] At present, there are many problems in the point cloud segmentation of complex and special-shaped metal workpieces in industrial scenarios. First, after the metal surface is irradiated by a laser area array camera, the surface is easy to reflect light, resulting in defective point cloud data collection, and it is impossible to obtain a complete point cloud; In complex industrial scenes, the point cloud segmentation of specific target objects has the problems of high environmental noise and complex background points that are difficult to directly segment; third, it is difficult to use functions to fit the surface of complex and special-shaped workpieces, and the commonly used random sampling consis...

AI Technical Summary

Problems solved by technology

[0002] At present, there are many problems in the point cloud segmentation of complex and special-shaped curved metal workpieces in industrial scenarios. First, the metal surface is easy to reflect light after being irradiated by a laser area array camera, resulting in defective point cloud data collection, and it is impossible to obtain a complete point cloud; In complex industrial scenes, the point cloud segmentation of specific target objects has the problems of high environmental noise and complex background points that are difficult to directly segment; the third is that it is difficult to use functions to fit the surface of complex and special-shaped workpieces, and the commonly used random sampling consistent point cloud segmentation The algorithm cannot accurately fit the special-shaped surface, and the segmentation effect is not ideal. Fourth, in the field of deep learning, the calculation amount of processing 3D point cloud data has a dimension increase compared with that of 2D images. The training time of the network is very long, and it is similar to the depth of PointNet. The neural network of the learning framework is difficult to extract the detailed features of the local point cloud

Fifth, there are very few data sets for complex surface point cloud data, and the neural network trained using the classic Shapenet data set has the problem that the segmentation effect of complex surface point cloud does not match

[0003] Aiming at the problem that the metal surface is easy to reflect light, when collecting the point cloud data of such a metal workpiece, the usual method is to coat the metal surface with a paint that is not easy to transmit light. However, this will cause the surface of the metal workpiece to need decontamination process and the thickness of the workpiece will change in the later stage. New problems such as the need for later data reprocessing arise; for the 3D measurement of aeroengine blades in an industrial environment, directly using the point cloud segmentation algorithm will easily generate a large number of background noise points, which will reduce the efficiency and accuracy of point cloud segmentation. A certain preprocessing method will eliminate the background points; for complex irregular surfaces such as aero-engine blades, the traditional RANSAC algorithm cannot accurately find an accurate function model that fits the engine blades, and cannot achieve accurate point cloud segmentation

[0004] Therefore, how to solve the low degree of automation of the current industrial measurement of complex special-shaped surfaces, and then realize the rapid and accurate segmentation of point clouds of special-shaped surfaces has become a problem that technicians in the field are dedicated to research.

Images