Research On Object Detection And Grasp Detection Based On Deep Learning

Currently,intelligent hand eye systems perform grasping tasks combined with de ep learning detection technology.The visual detection network based on deep learning has the ability to achieve a high accuracy.But the enormous network model and comp lex algorithm require a more high-end graphics card otherwise the detection speed wil 1 be damaged by a large amount.firstly,this paper focus on lightweighting the existing detection network in order to speed up the detection speed and reduce the need for hig h-end graphics card.secondly,this paper designs an multi-modal input network that ca n solve the random grasping pose problem.the main contributions are as follows:In this paper,convolutional neural network is applied to perceive the workspace.By pruning the network and optimizing the network structure,this paper achieves the purpose of accelerating the detection speed of the network and reducing the model co mplexity of the network.The number of parameters is reduced from 60m to 5m,the m aximum speed can reach 312fps,the accuracy on VOC dataset reaches 77.8mAP.A generative symmetry network for pixel level prediction of the workspace is des igned by using the combination of deep learning and depth camera.The network can i ntelligently detect the grasping posture of an object,and the network model is lightwig ht.the number of parameters is only 1.4 million,which is smaller than the model para meter of existing networks.Moreover,the network based on generation has a great im provement in the running speed,it can reach the running speed of 16ms per frame.Th e model achieved 98.8%accuracy on Cornell grasping dataset.