Research On Robot Workpiece Location Technology Based On 3D Model Matching

Against the backdrop of a fading demographic dividend and rampant epidemics,labour costs in industrial production are rising year on year.This has placed higher demands on the level of intelligence and automation in industrial production.In the field of intelligent manufacturing in industrial production,assembly robots play an unparalleled role as the core equipment of flexible automated assembly systems.Thanks to the rapid development of deep learning and computer vision technology,vision systems have become an important part of the unmanned operation of assembly robots.The use of vision-equipped assembly robots can significantly reduce labour costs and increase productivity.In vision system applications,there are three main approaches to artefact localisation: image feature-based,point cloud-based and deep learning-based.The image feature-based approach is more texture-dependent and is highly susceptible to lighting conditions.The point cloud-based approach,on the other hand,requires the construction of extremely accurate point cloud models.Using deep learning methods,this paper investigates 3D model matching and positioning techniques for robotic workpieces,with the main work including:(1)The study developed a workpiece positioning system that uses neural networks for 2D-3D keypoint matching in order to ensure both real-time and accuracy.A Pn P solver is then used to solve the workpiece poses from the key points.In addition,the software architecture and the hardware platform architecture of the system are described.(2)For the purpose of positional training,the image acquisition process in this paper uses a Lucid Helios + Triton combination industrial camera in combination with a non-self-AR toolbox such as ARkit or ARcore to acquire colour images containing time series,and relative camera poses,and then extract the feature points on the 2D images to construct a 3D sparse point cloud of the workpiece by means of Sf M.(3)To ensure the accuracy of the positional calculation,2D-3D keypoint matching was selected as the main strategy;an attention mechanism was introduced in the process of building the graph neural network model,and the matching prediction between 2D-3D keypoints was achieved through training.(4)From the 2D-3D matching prediction results,the Softmax operator is used to extract the matching confidence,and the RANSAC algorithm is used for the problem of key point matching solution.The solution results are further solved by the EPn P method to produce the final localisation results of the artefacts,and the visualisation of the localisation results is performed using a directed wraparound box as a medium.(5)Experiments were designed to test the dataset and demonstrate its validity.A comprehensive experiment was designed to evaluate the positioning results by introducing ADD metric pairs as the standard,and the results showed that the number of posture estimation trials with an error of less than 10 mm accounted for 89.87% of the total number of trials.To further validate the accuracy of the system’s estimation results,a six-degree-of-freedom platform robot was used to carry the workpiece in motion,and the posture feedback from the platform was compared with the visual positioning results.Finally,unconventional conditions such as occlusion,local information loss and complex backgrounds were set up to verify that the system is resistant to interference.In general,the construction of a vision-based workpiece positioning system for robots has been completed,which guarantees a certain degree of robustness under complex working conditions while achieving a certain level of accuracy.