Research On Grab Detection Method Based On Transfer Learning

With the rapid development of AI,the traditional grasp method of manipulator is becoming more and more intelligent.However,it is still a challenge to endow the manipulator with the ability to grasp with high precision like a human hand.It is an important ability to measure the manipulator to classify and generate grasping position parameters autonomously,which is also the focus of scholars in domestic and abroad in this field.In the environment of multi object stacking,the task of recognizing the target position and generating the grasping parameters also puts forward higher requirements for the manipulator grasping.In the present paper,the whole detection task is divided into target recognition classification task and capture detection task.The target recognition network is applied to the detection network by using transfer learning,and the influence of region level detection and pixel level detection on the detection accuracy is studied respectively.The main contents of this paper are as follows:(1)In order to obtain high-quality target feature information and reduce the difficulty of subsequent generation of grabbing parameters,the present paper uses transfer learning method to build target recognition network to extract and classify target objects.The small datasets are labeled again,and the datasets are preprocessed and divided into corresponding proportion of training set,verification set and test set.The designed network is trained through the training set,and the network accuracy is tested through the verification set and the test set.The test results show that the transfer learning method can effectively improve the accuracy of feature extraction.(2)For the sake of improving the accuracy of grasping parameters,a detection network based on region recommendation is built and studied.Using the transfer learning method,the constructed target recognition network is used as the feature extraction network of the detection network.The grab detection network uses the improved anchor box mechanism and related optimization algorithm to improve the quality of the generated grab parameters.In the present paper,small datasets are used for training verification and testing,and the accuracy of grasping position prediction reaches 95.32%,which meets the accuracy requirements.(3)For the purpose of improving the generalization of grabbing representation and the quality of grabbing parameters,the present paper studies the pixel level detection method and builds the corresponding pixel level detection network.In the present paper,a general-purpose representation method is proposed to meet the parameter representation requirements of different grabbers.At the same time,a grabbing detection network including feature extraction is built based on the idea of transfer learning and semantic segmentation.The new representation is used to label the small data set again,and the network is randomly divided into training verification set and test set to train the network.The accuracy of grasping position prediction is 93.47%,which meets the accuracy requirements.