Target Detection And Pose Esimating For Grasp Based On Deep Learning

Since the 1960s,robots have been introduced into industrial production with the advantage of being able to work continuously for 24 hours and adapting to the environment.With China's entry into the information age,robots have played an increasingly important role in various fields such as industrial manufacturing,family life,and national defense security.At the same time,they have put forward higher requirements for accuracy,real-time,and high efficiency.As an important part of robotic environment perception,machine vision has become a hotspot and frontier of robot perception research.Target detection and grab position estimation are the basic capabilities that industrial robots should possess.They play a vital role in industrial sorting,palletizing,assembly and so on.However,the traditional target detection and grab position estimation methods mainly rely on manual extraction of features and require a lot of experience accumulation for the important part of parameter adjustment.A lot of energy is wasted in repeated physical labor,professional talents are missing and detection accuracy is not high.The deep learning technology extracts features through the neural network,and under the guarantee of big data and computing power,the detection speed and detection accuracy can be greatly improved.This thesis aims to further improve the detection speed and accuracy by studying the target detection and object capture algorithms based on deep learning,and accelerate the formation of related technology applications.Secondly,this thesis studies the research method based on deep learning.Through the learning and understanding of Cornell's data collection,a data set for the target to be detected in a specific scenario was created.In the original grab detection network,the cavity convolution layer is added,the receptive field of the feature map is enlarged,and the features are further extracted.The deep and shallow feature maps are spliced to realize feature fusion,and the feature expression ability is further improved,so that the detection is performed.The accuracy rate has been significantly improved.Considering the accuracy and stability of industrial applications,a method based on corner detection is proposed for the target to be detected.In the scene,the object is classified and positioned first,then the corner point detection of the object is performed,and the detection point of the object is filtered and misdetected,and the corresponding algorithm is designed to complete the corner of the missed detection,so that the accuracy of the corner detection is greatly improved.Improve,and finally use the position information of the corner point to calculate the centroid position of the object,that is,the grab point of the object.The point cloud information is acquired by the depth camera,and the grasping posture of the object is calculated.This method can greatly improve the accuracy of object capture detection for specific scenes.Finally,the hand-eye calibration of the robot arm is carried out.The detected position information is sent to the robot through the ROS system.The robot grabs the object according to the instruction to achieve a simple object sorting scene.This thesis has important value and significance for the application and development of deep learning technology,as well as practical applications such as intelligent manufacturing,video surveillance,and automatic driving.