Research On Tracking Registration Of Mechanical Product Assembly Augmented Reality Based On Deep Learning

As a crucial link in integrated manufacturing,mechanical products' assembly has high technical requirements,long time consuming,and complex installation conditions.The assembly quality and efficiency of mechanical products directly affect their performance and production cycle.At present,in addition to the assembly of a small number of mechanical products that can be directly realized by using product threedimensional design information to realize automatic assembly,most assembly work is still guided by on-site assembly personnel according to the assembly manual.However,there are some problems such as low efficiency,low pass rate,and low quality in manual searching for assembly position according to assembly manual.In recent years,with Augmented Reality(AR)technology development,a new technical solution has been provided for the assembly of mechanical products.Augmented Reality technology can significantly reduce operators' cognitive and memory burden and improve the quality and efficiency of the assembly operation.As a crucial link in the process of integrated manufacturing,mechanical product assembly has high technological requirements,long time consuming,and complex installation conditions.The real-time and robust 3D tracking registration technology of machinery is the key to assisting the AR assembly system in realizing real-time and accurate superposition of virtual information in a natural assembly environment.In recent years,deep learning has been booming,and how to apply deep learning to 3D tracking registration has aroused extensive research among scholars.Aiming at the shortcomings of the traditional tracking registration method,such as inaccuracy and robustness when processing mechanical product assembly,this paper proposes a real-time,accurate,and robust deep learning 3D tracking registration method.The main work of this paper includes the following aspects:1.Given the scarcity of 3D data sets based on deep learning,complex labeling,and lack of universality,this paper conducts simple modeling for the target mechanical product through the depth camera.It generates 6D attitude estimation data set and object label annotation in the natural environment.To enhance the data set's generalization ability,this paper selects different background images of the SUN2012 data set and composes the 6D attitude estimation data set under the virtual environment of the target mechanical product in the Blender virtual environment.The experimental results show that the 3D data set produced by this method has a good tracking effect and good generalization ability.2.Monocular image of 3D tracking registered research focus has been enhanced assembly,this paper through the Farthest Point Sampling(Farthest Point from,FPS)algorithm to select points on the target object Point cloud,and then use the 3D detection based on pixel vote online to instances of output image segmentation and pixel Point to Point a unit vector,and then use Random Sampling consistency(the Random Sample Consensus,RANSAC)algorithm to reduce pixels vote by mistake,Finally,the pose of the object is solved by the uncertainty-driven Perspective-N-Point(PNP)algorithm.3.Aiming at the problem that the number of parameters and the amount of computation of the existing 3D detection network based on pixel voting is too large,this paper introduces the GHOST module in the lightweight neural network Ghost Net to compress the model.To compensate for the decrease in pose estimation accuracy for the lightweight network,the Squeeze and congestion(SE)and loss function are both there to improve the model accuracy.The experimental results show that the improved algorithm in this thesis retains the original performance while the number of parameters and computation is significantly reduced.4.Based on the algorithm in this paper,the existing AR assembly guidance system in the research group was tested through a specific type of semi-steel cable assembly.The user experience was investigated and analyzed,which verified the practicability and reliability of the algorithm.