Font Size: a A A

Part Recognition And Pose Estimation Based On Convolutional Neural Network

Posted on:2022-06-05Degree:MasterType:Thesis
Country:ChinaCandidate:C M LiFull Text:PDF
GTID:2518306566461304Subject:Mechanical engineering
Abstract/Summary:PDF Full Text Request
With the popularization of industrial robot applications,the automation level of industrial production continues to improve,and automated assembly and sorting based on machine vision have important application prospects in automated industrial production.Compared with traditional manual assembly,automated assembly based on industrial robots has significant advantages in terms of efficiency and stability.Recognition and pose estimation of parts in industrial scenarios is of great significance in improving the automation of robot assembly and sorting.Aiming at the problems of low part recognition efficiency,low pose estimation accuracy,and difficulty in obtaining a large number of training sets in industrial environments,the part recognition and pose estimation based on deep learning are studied,and a new method based on convolutional neural network and virtual depth image is proposed.Part recognition and pose estimation method of data set.First,use OSG(Open Scene Graph)to build a virtual training set to solve the problem of difficulty in obtaining and labeling a large number of training sets,and then use convolutional neural networks to realize part recognition and pose estimation in industrial scenes.The main tasks completed are as follows:(1)Constructed depth image data sets for part recognition and pose estimation,including virtual data sets generated by OSG and real data sets shot by depth cameras.The OSG and physics engine are used to render a large number of virtual depth image data sets of the cluttered scene in the virtual scene,and the virtual depth image data sets of the part poses obtained by rendering various parts at different pose angles.The real data sets is obtained by using Kinect 2.0 depth camera to collect the depth image of the part and repair it.(2)Part recognition method based on convolutional neural network and virtual training set is studied.Convolutional neural network models with different depths and structures are constructed and trained with virtual depth images.After the training is completed,real data sets are used to test the network model and select the part classification model with higher recognition accuracy.Experiments show that the method is trained by virtual training set,and it can recognize the real parts accurately.At the same time,the depth image is robust to the reflection and less texture of the parts.(3)The pose estimation method based on FSA-Net is studied.The head pose estimation network FSA-Net is applied to the pose estimation of industrial parts.Firstly,the algorithm structure of FSA-Net is analyzed.Then,the influence of feature scoring mechanism and fine-grained structure mapping on the network framework is analyzed through experiments.By controlling the number of training sets and ensuring the Angle of pose estimation,the pose estimation of industrial parts is realized accurately.On the virtual test set,the average error of the part pose estimation is 1.92°,and on the real test set,the average error of the part pose estimation is 6.31°.(4)A Fully Convolutional neural network(FCN)based part segmentation method for cluttered scenes is studied.A full convolutional network model is constructed for semantic segmentation of cluttered scenes,and pixel-level classification of parts is realized.In the training,the loss function of the model is quickly converged by adjusting the hyperparameters.After the training is completed,the accuracy of the part segmentation of the network model in the cluttered scene is tested.The experiment proves that the accuracy of the part segmentation pixel on the virtual test set reaches91.42%.The segmentation pixel accuracy of the parts on the set reaches 85.66%.(5)Through the integration of the cluttered scene segmentation module and the part pose estimation module,a network framework for the pose estimation of the cluttered scene parts is built.First,the parts in the scene are segmented,the depth image is extracted using the segmentation results,and then the positioning and pose estimation are performed.After the training is completed,the virtual test set and the real test set are used to test the parts recognition and pose estimation accuracy of the network in the cluttered scene.Experiments have proved that the pose estimation method for parts in a cluttered scene proposed in this paper can accurately segment and estimate the pose of parts in the scene,and provide a stable information basis for industrial automated sorting and assembly.(6)Part pose estimation method based on PVNet(Pixel-wise Voting Network)in cluttered scene is studied.Firstly,the depth images and RGB images of cluttered scenes are collected by the depth camera,and the transformation relationship between the parts and the camera coordinates is marked by the camera calibration plate.Then,the collected images are used to construct the point cloud model,and the segmentation mask and pose label are generated by the point cloud model.After the preparation of the training set,the part pose estimation in the cluttered scene is realized by training and testing the PVNet network.
Keywords/Search Tags:convolutional neural network, virtual data set, depth image, part recognition, part pose estimation in cluttered scene
PDF Full Text Request
Related items