Research On Detection Algorithm Of Robotic Arm Grasping Based On 2D And 3D Vision

With the proposal and implementation of major national strategies such as "German Industry 4.0","Japan’s Revitalization of Manufacturing Industry" and "Made in China 2025",the world is experiencing a period of great changes in the rapid development of robots.Production applications are becoming more and more extensive.In this paper,the scattered and stacked targets are taken as the grasping target,and the industrial corner pieces with complex shapes are selected as the grasping workpieces,and a set of grasping system based on 2D/3D vision is built.6D pose estimation method,and in order to improve the workpiece recognition rate and grasping efficiency,the linear push strategy is studied,and the push strategy is applied to the grasping system.The main research work is as follows:First of all,the needs and difficulties of robots in the field of industrial disordered grasping are deeply investigated,the research status of grasping detection based on 2D/3D visual data at home and abroad is analyzed,and the functional characteristics required for practical application of robot grasping are analyzed.Conclusion:(1)The early configuration process is simple and fast;(2)The types of workpieces to be grasped are diverse;(3)High recognition speed and grasping efficiency;Hardware selection is carried out according to the needs,and a grasping system based on 2D/3D vision is built.Secondly,based on the perspective projection theory,the camera imaging model and projector imaging model are analyzed,and the internal and external parameters of the camera,the TCP calibration of the robotic arm,and the handeye calibration are performed.Take the physical simulation platform to study the acquisition of simulated data and compare it with the actual data;based on the Qt application development framework,build the software system of the grasping platform,including the communication between the robot arm,the camera and the end effector,synchronous control,algorithm embedding,real-time Simulation and other functions.Thirdly,research the target 6D pose estimation algorithm,analyze and apply the PPF template matching algorithm,build the 1.0 version of the grasping system,and realize the basic grasping function,but the experiment shows that the algorithm will have a large mismatch rate;The PPR-Net deep learning algorithm is used and analyzed,and the CAD model and physical simulation technology are used to construct a training dataset without manual annotation,and the generation includes RGB images and depth images in different stacking situations,as well as the corresponding workpiece category,visibility,6D bit The pose label is trained in the deep learning network,and the pose estimation test experiment is carried out.The experimental results show that the number of correctly estimated workpieces based on the PPR-Net deep learning algorithm is twice that of the PPF template matching algorithm,with higher accuracy;Research the grasping detection algorithm combining RGB image(2D)and point cloud data(3D),and propose an improved PPR-Net deep learning algorithm based on 2D/3D vision,which effectively improves the grasping efficiency.Finally,aiming at improving the recognition accuracy of the robot in the small workpiece grasping scene,on the basis of the development of the existing2D/3D sensor,the deep features of the RGB image and the 3D point cloud are integrated,and an adaptive clustering grid is proposed.The algorithm is based on the method based on the algorithm construction,and integrates the point cloud height rendering information to realize the algorithm strategy of the toggle path planning.This strategy can improve the average target matching recognition rate to 39.6% in the simulation scene and 16.4% in the actual scene.Verify the effectiveness of the algorithm.