Research On The Construction Of Virtual Datasets And Intelligent Recognition Method For Bin Picking

Bin Picking is a typical problem in identifying and sorting complex stacked objects.Aiming at the difficult problems such as the serious stacking of target objects in Bin Picking and the weak robustness of traditional detection algorithms,this paper uses the method of deep learning to realize the instance segmentation under this scene,deeply studies and constructs the deep learning framework of instance segmentation,so as to lay a foundation for the subsequent object grabbing.In view of the difficulty of obtaining target datasets in deep learning,this paper adopts virtual simulation technology to construct virtual datasets,data augmentation and other methods to expand the training dataset,and realizes the effective training of deep learning model.The main research contents of this paper are as follows:(1)On the basic of virtual simulation technology,the design of virtual dataset construction schema is put forward.The construction schema uses OpenGL and Bullet physics engines to construct the scene of object stacking,using the method of viewport rendering to generate virtual datasets,thus avoiding the tedious work of manually making datasets.(2)This paper studies the data augmentation method of dataset,puts forward the method of data augmentation pipeline in an innovative way,guarantees the accuracy of the dataset after data augmentation through the method of connected domain processing,and improves the efficiency of data augmentation processing by using multi-thread and multi-process method.(3)This paper studies the theoretical basis of deep learning.By using the top-down method,the overall task is divided into four sub-tasks of feature extraction with convolutional neural network,region proposal in the image,region classification and border positioning,and pixel level segmentation in the region.On this basis,this paper gives the training strategy of the deep learning model.This paper designs and carries out the experiments of virtual dataset generation,data augmentation and intelligent recognition.The experimental results prove that the deep learning model established in this paper has a good recognition effect with the use of virtual datasets and data augmentation,so it provides a good application basis for the target recognition of Bin Picking.