Research On 6D Pose Measurement Algorithm Based On Virtual Simulation Of CAD Model

With the rapid development of computer,electronic communication and other technologies,machine vision has been widely used in various industries.Among them,the 6D pose measurement method based on CAD model has a high measurement accuracy,which can well match and apply to intelligent scene operations such as visual-based robot grasping and assembly,but it also has the problems of time-consuming and unstable model base matching.In recent years,the emergence of deep learning has broken through the limitations of traditional image processing algorithms.It can learn,judge and recognize images without relying on artificially designed features and templates,and has strong stability and high flexibility.Therefore,the development and application of deep learning can effectively make up for the shortcomings of traditional image processing algorithms.However,the problem is that the training of an effective deep learning network requires large-scale data,and large-scale data collection and labeling are time-consuming and labor-consuming,which cannot be applied to the requirements of high efficiency and cost control in industrial production.This has also become the primary factor limiting the application scope of deep learning in industry.Therefore,in view of the above problems,this paper combines the traditional algorithm with deep learning method,uses the powerful feature extraction ability of deep learning to conduct rough positioning of the detected target,and then uses the traditional algorithm to conduct 6D pose measurement in the region of interest.In this paper,the data,methods and actual effects used before and after the improvement are compared and analyzed,and the following work is mainly done:(1)Data generation method of virtual simulation based on CAD modelIn order to solve the problem of time-consuming and labor-consuming in the process of industrial site platform construction and data acquisition,this study designed a set of method of parameter-generated virtual simulation workpiece image data set based on the CAD model commonly used in industrial production and the actual industrial application background.By acquiring the CAD model of the workpiece to be detected,the texture and material of the workpiece are simulated,and then imported into the OpenGL graphics library for environment simulation(including all kinds of light sources and industrial background,etc.).The perspective of the virtual camera and the pose of the workpiece model are adjusted,and the sample is carried out in a certain space.This method can obtain a large number of training samples suitable for deep learning target detection in a short time.Finally,Labelme software is used for unified labeling and data sets are divided.The virtual simulation data and the field data are compared and analyzed.(2)Deep learning target detection based on virtual simulation dataIn order to solve the problem of time-consuming and unstable template matching process in 6D pose measurement algorithm,this paper proposed an improved deep learning network based on YOLOV3 for industrial parts target detection,which used lightweight workpiece images established by virtual simulation method for training instead of actual images collected on site.At the same time,the model and parameters are improved and optimized to improve the performance of deep learning network.Experiments compare the performance of the YOLOV3 network on the virtual simulation data set before and after the improvement,and make a specific analysis.The improved optimal model is saved and tested and analyzed on the virtual simulation data set and the actual workpiece pictures.(3)Real time stable 6D pose measurement system based on deep learning network and CAD modelIn this paper,a feasible 6D pose measurement system is designed.Based on the analysis of actual needs in industrial applications,functional design is carried out with the combination of deep learning method,mainly including rough positioning module and pose measurement module,which can display and record the positioning and measurement process,and save the test results.Finally,the system is tested under different conditions,showing the practical value of the system.