Research On Robot Target Recognition And Grasping Based On Convolutional Neural Network

Robot capture technology has become an important direction in the field of robot research.Although many scholars have invested a lot of energy in robot capture,the problems of inaccurate recognition of grab points and poor real-time performance still exist.Aiming at this problem,this thesis designs a grab detection algorithm based on RP-ResNet network,and puts forward a multi-task convolution neural network model considering the actual grab needs,which can perform two tasks: grab detection and target classification at the same time.We set up a robot grasping experiment platform,the real-time,accuracy and feasibility of the robot grab network are verified.(1)Firstly,the imaging principle and calibration method of binocular camera are analyzed,then the camera parameters are calibrated by Zhang's calibration method,and the camera calibration is realized by using Matlab software.Finally,the relationship between camera parameters and camera position is obtained.(2)This thesis designs a convolution network model RP-ResNet,which is based on ResNet-50,makes full use of the features of low-level position feature information of deep neural network and rich deep-level grasping features,adds regional recommendation network(RPN)to the low-level network,and extracts the first stage of the input image.In the second stage,an improved pyramid pooling structure is adopted to pool the feature map with different resolutions,so as to retain more original image information,increase the data dimension of output,and use SENet network to enhance the attention of the channel,thus enhancing the detection accuracy of the network to small target objects.Using the augmented Cornell data set divided according to different ways to train the RP-ResNet network,the accuracy of the test results on the test set is as high as 96.72%,and the detection time of grasping is reduced to 0.12 s.(3)The multi-task output is realized by using the proposed multi-task convolution neural network.According to the parameter sharing mechanism of the multi-task convolution neural network,the cross-stitch module shares the information between the main task grab network and the auxiliary task classification network.The traditional Inception network is combined with the Residual module,and the Reduction module is used to replace the traditional pooling layer to keep the classified information to the maximum extent,which realizes the mutual promotion of the chief and the auxiliary tasks and the mutual utilization of the characteristic information.(4)A binocular vision robot grab system is built,and the Six-Degree-of-Freedom manipulator is modeled and simulated by Matlab,which verifies the feasibility of the positive and inverse solutions of the manipulator.By using ABB Robot,the robot grabbing experiment is carried out,and the success rate of object grasping reaches 86%,which verifies the feasibility of robot grabbing system.