Research On Grasping Positioning Technology Based On Kinect Camera

In recent years,robot grasping technology has been widely used in all walks of life in social production and life.As the application scenarios of robots become more and more extensive,the ability of robots to analyze scene information is required to become stronger,so accurate grasping of any object and any posture in a complex environment has become a hot research direction of robot grasping technology.This thesis focuses on the research of grasping positioning algorithm,aiming to determine a more accurate grasping pose to improve the grasping success rate.In this thesis,the robot arm is integrated with vision technology and deep learning.Among them,the Kinect camera is used to enhance the environmental perception of the robotic arm,and the deep learning method is used to obtain a more robust grasping pose.Finally,an experimental platform for object perception?grasping positioning and grasping execution is built.In the object perception link,this thesis uses the Kinect camera to obtain RGB-D images,and proposes a method that combines pixel filling and an improved joint bilateral filtering algorithm to repair the depth map holes.In the target capture and location link,for different target scenarios,the two deep learning-based target capture and location algorithms proposed in this thesis are as follows:1)For single-target scenarios,Drawing lessons from the capture and positioning network structure based on region extraction,study the advantages and disadvantages of the dense module and the residual module to improve the original network.This thesis designs a grasping and positioning model that uses Dense Net-121 network instead of Res Net-50 as the feature extraction structure.In view of the disadvantages of the original network model in the performance of small target detection,the feature pyramid was added in the feature extraction process,and ROI Align was used to replace the network structure of ROI Pooling,which solved the multi-scale problem in grabbing positioning and inaccurate grabbing frame positioning.The problem.Finally,the idea of grid division is proposed to improve the grasping suggestion network,which reduces the repeated estimation of the area and balances the detection speed and detection accuracy.2)For multi-target scenarios,the target detection algorithm and the grasping positioning algorithm are integrated.In order to meet the requirements of high detection accuracy and high detection speed,both algorithms use the YOLOv4 network structure.In the grasping and positioning algorithm,the SPP module is firstly added to improve it;secondly,the multi-proportional rotating anchor frame mechanism adapted to the grasping characteristics of the robotic arm is used to predict the grid;the final experiment shows The combination of YOLOv4 and the rotating anchor frame is more robust to the positioning of objects with multiple grasping methods.In the grasping execution link,the “eye out of the hand” calibration method is used to obtain the spatial coordinates of the graspable position of the target object,and the linear interpolation algorithm in Cartesian space is used to plan the trajectory of the manipulator.The physical platform is built to complete the positioning and grasping tasks of single and multiple targets.The accuracy of single target positioning and grasping detection is 92.5%,and the grasping success rate is 86.5%.For the multi-target positioning and grasping experiment,first use the target detection to segment into a single object,and then use the grasping positioning technology to determine the grasping pose.The grasping detection accuracy rate is 95.5%,and the grasping success rate is 92.7%.It verifies the feasibility and robustness of the target grabbing and positioning algorithm proposed in this thesis.