Research On Visual Positioning Technology Of Workpieces Based On Prediction Of Key Points

As the aging of the population is increasing year by year and labor costs are gradually increasing,the industrial production process is developing in the direction of intelligence and automation.Robots are very common to replace manual production.Our country's 14 th Five-Year Plan and 2035 The long-term goal outline clearly includes robotics as a key breakthrough scientific and technological field,and other countries have also issued relevant strategies to promote the development of robotics.As a key process in the product production processes,the assembly process accounts for about 50% of the total production time and more than 30% of the total cost.The usage of assembly robots for production operations can greatly improve production efficiency and reduce costs,which is of great significance.In recent years,the rapid development of artificial intelligence and computer vision technology has provided a direction for robots to achieve a higher degree of intelligence and automation.Many robot products today are equipped with vision systems that give robots the ability to perceive themselves and enable unmanned operations.Among the lots of visual detection methods,according to the different detection features,it can be divided into three categories: manual features,deep learning,and point cloud features.Hand-based features are easily affected by interference factors such as lighting conditions,background,and occlusions.While the use of point cloud features requires higher accuracy in point cloud model construction.In this paper,we use deep learning method,according to actual problems,researching on the visual positioning technology of workpieces based on six degrees of freedom pose estimation.The main work is as follows:(1)A visual positioning system for workpieces was built,which first predicted the vector field of key points on the workpiece by a deep neural network.Then the key points are voted by the vector field.And finally the pose of the workpiece are calculated from the key points.The hardware composition and software design ideas of the system are discussed.(2)Aiming at the problem that the real poses are difficult to obtain during the preparation of the pose estimation training data sets,this paper uses the Real Sense F200 depth camera as the image acquisition device,combined with the Ar Uco pose detection logo and ICP point cloud registration technology.And we use such an offline calculating method,the pose information of the workpiece is successfully calculated,and a data set containing color images,depth images,workpiece mask images,workpiece poses and other information is constructed.(3)We use the FPS algorithm as the main strategy to select the target key points from the workpiece surface.Then we use the PVNet deep network framework to build a deep neural network model,and conduct the semantic segmentations of the workpiece area in the input images and the prediction of vector fields for each key point after training.(4)The key point hypotheses are generated from the vector field prediction result.And the voting points voted on the key point hypotheses according to the voting strategy based on the RANSAC algorithm.The voting results are analyzed,and the voting strategy is improved and optimized.According to the distribution of the voting results of the key points,the reference points are determined.And the Pn P solver calculated the pose of the workpiece.(5)Through the data set verification experiment,the validity of the data set is proved.The pose estimation experiments are carried out,and the workpiece pose positioning result is evaluated according to the ADD metric,and the accuracy rate of the pose estimation result is 86.05%.The Steward platform drives the movement of the workpiece,the poses of the visual positioning system and the platform are compared,which verifies the accuracy of the system's estimation results.Through experiments under the conditions that the part of subject is blocked,cut off,or with the complex background interference,human interference,etc.,the system is verified to have a certain degree of robustness.