Research And Application Of Robot Target Detection And Grasping Detection Based On Deep Learnin

Robot grasping technology is one of the important research directions in the field of robotics.At present,most mature robot grasping technologies are only applicable to the structured operating environment with relatively fixed layout.When there are a variety of objects with rich features and arbitrary poses in the grasping scene,it is difficult for the robot to effectively recognize and grasp the objects in such a complex environment,and the grasping real-time performance is poor.In order to improve the success rate and efficiency of robot grasping in complex environments,the first thing that needs to be improved is the robot’s perception ability of the complex external environment.In this paper,some key technologies of visual perception involved in the robot grasping process are studied,mainly including target recognition and positioning technology,optimal grasping pose detection technology and camera calibration technology.Firstly,the robot grasping system is designed and analyzed.The hardware implementation scheme of the system is elaborated,with the research of depth image hole filling.The transformation matrix between the robot and the visual perception system is obtained through camera calibration,and the forward and inverse kinematics of the manipulator are analyzed.Secondly,aiming at the problem that the traditional target detection methods perform poorly in the recognition and location of the objects to be grasped in complex grasping environments,a detection algorithm of the target to be grasped with a rotated bounding box is designed.The improvements of the YOLOv5 target detection algorithm are carried out with the referring to the Ghost Net network.The C3 module in the backbone network of the YOLOv5 model is replaced with the redesigned C3-Ghost Bottleneck module that is more efficient with fewer parameters,and some regular convolutions in YOLOv5 are replaced with Ghost convolutions,further reducing the count of parameters in the model.By adding the prediction branch of Angle to the detection layer of the YOLOv5 model,and introducing the Densely Coded Labels(DCL)technology,the detection algorithm YOLOv5-Ghost-Angle with rotated bounding boxes is obtained.The experiment is carried out on the self-made grasping dataset,and it is verified that the detection accuracy and speed of the improved model are better.Then,to address the problem that current mainstream grasping detection algorithms have difficulty balancing detection accuracy and speed at the same time,a new grasping detection network is designed.The lightweight network Rep VGG is applied to grasping detection for the first time and integrated with the Efficient Channel Attention(ECA)mechanism to obtain a more efficient feature extraction backbone network,Repv GGECA.The oriented anchor mechanism is introduced,and its hyperparameters are set through clustering algorithm,and the matching mechanism is optimized.The model is trained and tested on Cornell dataset,achieving a detection accuracy of 97.4% and a detection speed of 25.45 FPS.Finally,the experiment is carried out on the real robot grasping platform.The various functional modules of the robot are connected by using Robot Operating System(ROS),and the grasping experiments are conducted on both single and densely arranged multiple targets,validating the feasibility of the overall algorithm in the actual grasping tasks and demonstrating its reliability for robot grasping in complex scenes.