| With the gradual popularization of industrial intellectualization,mechanical arms are widely used in the market as actuators that can replace manual operations.The control methods of traditional mechanical arms are mostly through manual teaching and offline programming.However,in the specific grasping work,there will be problems such as random posture,irregular shape,and complex environment of grasping objects.It is particularly important to effectively use visual perception to improve the grasping accuracy and real-time performance of the mechanical arm.Therefore,this thesis proposes a mechanical arm intelligent grasping algorithm based on deep learning,and designs a grasping method based on Convolutional Neural Network for Multiple Residual Extraction(CRE-Net)for plane 2D pose grasping.The algorithm solves the problems of grasping failure caused by the physical properties of objects and the influence of external environment in the traditional method.The details of the research are as follows:(1)In view of camera imaging distortion existing in the visual capture system,Zhang Zhengyou calibration algorithm is used in ROS to calibrate the camera and obtain the internal parameter matrix of the camera.Then,two methods of hand-eye calibration are studied and analyzed.And use AR tags in ROS to complete hand eye calibration and determine the camera’s external parameter matrix.(2)A pixel-based data re-labeling method is adopted to solve the problems such as the small amount of Cornell dataset and the lack of label data information.Building a dataset collection system using depth cameras to independently is to expand the amount of data captured in Cornell dataset,and uniformly pixel level re-label the expanded dataset to generate captured tags that can be used for two finger gripper grasping.This will solve the problems of small amount of labels,large occupied area of labels and poor accuracy of labeling in rectangular labeling.(3)This thesis proposes a grasping detection algorithm based on CRE-Net network to solve the grasping problems of mechanical arms such as random posture,irregular shape,and complex grasping environment.This thesis analyzes the construction process of the CRE-Net network and conduct research and analysis on its basic network composition module,process the algorithm based on the four grasping parameter information output by the CRE-Net network,and solve the grasping pose.This method can achieve stable grasping under different objects and environments,and solve the influence of external factors on the grasping robustness.Analyze its network evaluation criteria and compare the CRE-Net network with more popular crawling networks for model comparison,CRE-Net network pose accuracy of this network reaches 93.99%,and the detection effect is better.(4)The method proposed in this thesis is verified by real-time capture in Pybullet environment.The D-H modeling and forward and inverse kinematics analysis of the structure of the Franka Panda mechanical arms are carried out,and a complete grasping detection platform is built based on mechanical arms.Analyze the crawling detection process of CRE-Net network in the Pybullet environment,and build a complete crawling verification platform based on the Franka Panda mechanical arms.Under the platform,the two-finger gripper and the mechanical arms body are used to carry out grasping experiments,And to address the issue of the impact of increasing the number of objects being grabbed on the success rate of grasping,real-time grasping tests were conducted on single object grasping and multi object grasping in the same scene.The experimental results show that in adversarial grabbing,the success rate of grasping a single object is 94.8%,and the success rate of group grasping of five objects is 76.7%,which has a high success rate of grasping. |
