Segmentation And Registration Of Point Cloud Model Of Disordered Workpiece

As depth sensors gradually become cheaper,3D point cloud processing technology has begun to be widely used in three-dimensional reconstruction,robot grasping,target recognition and other fields.The 3D depth sensor based on line laser/structured light can acquire the point cloud of target objects in complex scenes without being affected by factors such as light,angle,surface texture,etc.It relies on data processing software to perform geometric description,positioning,guidance,measurement and inspection applications for products.With the implementation of the 2025 national strategy for smart manufacturing,machine substitution and smart factories have begun to be implemented in large-scale enterprises.The parts loading and unloading process urgently needs to use 6-axis robots to grab and unload materials to replace manpower.In order to achieve this goal,the Bin-picking system based on machine vision has attracted widespread attention and has gradually become a research hotspot.The difficulty of Bin-picking is that the workpieces are placed out of order and the workpieces block each other,which leads to problems in target recognition.In order to solve this problem,this paper uses the line laser scanning system to quickly obtain the part point cloud,and then segmented and registered with the standard template point cloud model.Finally,the obtained accurate 6D pose parameters are passed to the robotic hand to achieve target capture.The specific work includes the following aspects:(1)In view of the point cloud noise problem caused by sensors and the environment,the method of statistical filtering combined with bilateral filtering is used to denoise.Bilateral filtering smooths the internal noise of the point cloud,and statistical filtering eliminates outliers at the edge of the point cloud.The method of normal differentiation combined with voxel grid down-sampling is proposed to effectively down-sample the high-density original point cloud,reducing the number of points in the point cloud while retaining effective details.For traditional super-voxel segmentation problems that require manual setting of a large number of initial values,a minimization method based on exchange and fusion is used to generate adaptive super-voxel resolution,and the segmentation at the boundary is clearer.Through the clustering method of Euclidean distance combined with local geometric features,the cluster growth of super voxels is carried out,and the individual part point cloud is segmented from the complete point cloud.(2)Aiming at the problems of traditional Iterative Closest Point(ICP)reliance on good initialization quality and easy to fall into local optimal solutions,a combination of fast global point cloud registration method and improved ICP algorithm is proposed.First,construct a fast point feature histogram of the point cloud data,and then alternately optimize the different components of the objective function to obtain the optimal initial transformation when the feature matching relationship is determined,and finally use the improved ICP algorithm to accurately perform the initial transformation results match.The experimental results show that the method in this paper has a great improvement in matching accuracy,and the speed is an order of magnitude improvement compared with traditional coarse and fine matching.(3)The software development based on open source platforms such as Point Cloud Library(PCL)and Open3 D.A QT-based point cloud model registration system is designed.Through the interface operation,the result display is more intuitive.