Research On Key Technologies Of Stereo Visual Recognition And Positioning For Wire Rolling And Tagging Robot

With the development of the national manufacturing industry,the steel industry has also received increasing attention.As one of the steel products,hot-rolled coils account for over 20% of China’s total steel production.Before using the coil wire bundled by the bundling machine,a label with information such as steel grade,production date,furnace number,weight,etc.should be hung on the bundling line to indicate "identity".At present,most of this operation at home and abroad is still limited to manual suspension,and the speed,efficiency,and worker safety issues during the work process are difficult to guarantee.Therefore,it is necessary to achieve automation and intelligence in the operation of hanging signs.This article improves the PointNet++algorithm model to achieve recognition and segmentation of wire and bundle,and then obtains the corresponding three-dimensional coordinates of wire and bundle.Combining the three-dimensional data,the position coordinates of the suspension point are calculated to achieve the positioning of the suspension point.The experimental results are verified using industrial cameras and robots in laboratory and factory environments,and the project requirements are met.The automation and intelligence of the label hanging process are achieved.The research content of this article includes:(1)Obtaining point cloud images of target objects.Using a 3D binocular camera to collect images of coil wires and preprocess the collected point cloud images to reduce the number of point clouds while ensuring that the point cloud features remain unchanged.(2)We conducted recognition and positioning experiments on wire and bundle based on PCL point cloud template matching algorithm and YOLOv5 algorithm,analyzed the shortcomings of the two methods,combined 2D deep learning with point cloud,and decided to use 3D deep learning technology to complete the research project.(3)Produce a dataset for hot-rolled coils.This article uses the Cloud Compare tool to label 600 point cloud images,which are divided into two categories:wire and bundled wire.The labels correspond to 0 and 1,and the training set and prediction set are divided proportionally.(4)Split the hot rolled coil into wire and bundle wires.This article is based on an improved PointNet++ algorithm model to train and process target objects.In terms of input data set,sampling model,FP layer channel setting and data visualization,improvements have been made,which are suitable for the research operation in this paper.(5)Calculate the three-dimensional spatial coordinates at the suspension point.Utilize open-source Visual Studio and Python for communication calls,and calculate the hanging point of the label based on the idea that there is a gap between the wire bundle and the wire,and the gap size is greater than the diameter of the hanging needle.(6)Conduct experimental verification of the suspension point position.Visualize the position using point cloud images,and confirm the coordinates of the suspension point by using robots to reach the designated suspension point.Conduct experiments in the laboratory and actual site to ensure the accuracy of the suspension point position coordinates.Finally,after experimental verification,the prediction time for a single photo was around 280 ms,and the accuracy of the algorithm’s output correct coordinates was around94%,proving the good performance of the algorithm model in object recognition and positioning of wire coils.