Research On Planar Grasp Detection Method Based On RGB-D Image

As labor costs increase year by year,in high-intensity or high-risk repetitive tasks,the use of intelligent robotic arms instead of labor meets the needs of the times.With the continuous in-depth research of related computer vision algorithms,the robot's perception of grasping targets has received technical and theoretical support.In this paper,the problem of 2D grasp detection in robot grasping is researched,using image processing and deep learning technology,combined with target detection algorithms to design a basic network structure for grasp detection,and on this basis,an accurate The real-time 2D grasp detection method achieves a high grasp detection accuracy rate.First,this article analyzes the task scenarios and task goals of the 2D grasping scene,and summarizes and briefly analyzes the datasets and deep learning techniques involved.The basic network structure of the grasp detection network is designed,including feature extractor and grasping detector,and a shallow grasp detection model is implemented based on the AlexNet network.Through the analysis of the experimental results and theories of the above methods,the idea of grasp detection based on the object detection framework is proposed,and the rasterization idea and angle classification design are introduced into the grasp detection model.At the same time,the rotation data enhancement strategy and Multi-modal input,and based on this experimental verification and analysis,proved the performance of the algorithm,and laid the foundation for the optimization of the grasp detection algorithm.Secondly,a new type of grasp detection network for 2D grasp detection is proposed,which can predict the five-dimensional grasping configuration through images in real time.Combining the ideas of rasterization and angle classification,an end-to-end convolutional neural network is designed based on CSPDarknet.And for the first time,the attention mechanism is introduced to the grasp detection task,so that the grasp detection network can pay more attention to the potential grasping area in the image,rather than the large-area background image.At the same time,an angle label smoothing strategy for grasping angle classification is proposed.Compared with the previous method,it can distinguish the optimal angle,suboptimal angle and wrong angle in angle classification,so that the network improves the tolerance of the angle prediction.Experiments have been carried out on the Cornell dataset and the Jacquard dataset.On the basis of satisfying real-time performance,the method have achieved the highest accuracy.Finally,in the simulation environment,the scene of the 2D grasp detection task is built,and the simulation of the method in this paper is realized based on the physics engine and the forward and inverse kinematics module provided by the CoppeliaSim platform.While further verifying the performance of the algorithm,the applicable conditions of the algorithm in this paper are clarified in specific applications.