| Aiming at the problem of repeated grasping of auxiliary parts in the process of laying rails,a grasping detection algorithm is designed based on convolutional neural network to realize stable and accurate automatic grasping of auxiliary parts by robotic arms instead of manual work.According to whether the distance between the camera and the target changes along the optical axis,grasping detection can be divided into fixed plane grasping detection and variable plane grasping detection.In the fixed plane grasping detection,aiming at the problem of low grasping prediction accuracy,the Mask-RCNN network is used to segment the optimal grasping area of the target,and then ellipse fitting is used to obtain a highprecision grasping description.At the same time,the weight of the network is realized through the depth separable convolution to accelerate the prediction,and the data set is self-made based on the idea of dense grabbing and network training is completed based on transfer learning.The average prediction speed of the algorithm is 12 FPS.The error of the prediction center point in the x and y directions and the angle error are within 5pixel,3pixel and 3° respectively.The prediction speed is fast and the accuracy is high.Aiming at the problem that multi-scale targets are difficult to detect in variable-plane grasping detection,a grasping detection algorithm combined with historical results is proposed.The Keypoint-RCNN is used to predict the three key points in the target grasping area,and the anchor frame network is designed to improve the robustness of multi-scale target detection.Combined with the historical detection results,the network input and output are coarsely positioned and verified,which further solves the scale problem and improves the prediction speed and accuracy of key points.Finally,the RGB and RGD datasets are self-made and combined with the two datasets to complete the keypoint network training.After the improvement,the average detection speed of the algorithm is increased from 6FPS to 9FPS,and the average error of key point positioning is only1.63 pixels.Finally,based on ROS,a robotic arm autonomous grasping detection platform is built,which combines target recognition,attitude calculation,trajectory planning and control to improve the portability of the system.Based on Gazebo,the simulation experiment of fixed plane grasping is carried out,and the grasping success rate is 95%.A real experimental platform was built to verify the variable plane grasping detection algorithm,and the grasping success rate reached 96.7%,verifying the feasibility of the algorithm. |
