Research On Object Recognition And Robotic Pose And Pose Detection In Unstructured Environment

Robot gripping technology is one of the main research directions of robotics.At present,most mature robot crawling technologies are only suitable for a structured environment with a fixed scene layout.In the face of application scenarios with many object features and random poses,it is difficult to grasp and has poor real-time performance.In order to improve the efficiency and accuracy of robot gripping in unstructured environments,this paper studies the key technologies of robot gripping,including target classification and positioning technology,optimal gripping point detection technology,and visual perception system calibration technology.The details are as follows:(1)In order to solve the problem that the current regional recommendation detection model takes more time,the YOLOv3 algorithm is proposed to be used in the crawling system,and the performance difference between the YOLOv3 algorithm and other detection models is analyzed and compared.Use multi-label and cross-scale fusion to perform feature extraction and position calculation on objects to achieve object classification and positioning.The final experiment shows that the average accuracy of the detection algorithm is 0.92 and the average detection speed is 0.26 s.Compared with 172 s of FRCN,the real-time performance of the model is improved on the basis of ensuring a high recognition rate.(2)In order to solve the accuracy problem of end-to-end grab detection,a grab mapping relationship based on the normal direction of the object surface is proposed,and a novel grab detection algorithm combining adaptive threshold segmentation and residual network is designed.The normal-based crawling mapping relationship is used to uniquely convert the two-dimensional crawling parameters to the three-dimensional space,the adaptive threshold segmentation is used to obtain the object contour and generate all the candidate region sets,and the residual network-based model is used for Learn the mapping relationship between image data features and optimal grab points.The experiment is completed by means of transfer learning.The results show that the detection speed of the algorithm reaches 0.29 s and the accuracy rate is 89.7%.(3)In order to solve the problem of lens distortion of the vision sensor,a calibration experiment was performed using the Zhang Zhengyou calibration algorithm on the MATLABR2014 a application software to obtain the internal and external parameters of a single camera.Finally,the two cameras were jointly calibrated to obtain the rotation relationship between the two cameras and Translation relationship.Experiments show that in the end all errors are within 0.25 pixels,and the calibration is reliable.(4)In order to verify the feasibility of the above algorithm in an unstructured environment,a grab hardware platform of V-REP was built,and object detection,grab detection,and camera calibration techniques were combined to classify and locate objects,and the optimal grab points were obtained.Position trajectory planning for the robotic arm to achieve grasping operations.The operation control interface built by QT is divided into two groups to verify the feasibility of the target recognition algorithm and the grab detection algorithm for unstructured environments.The experimental results show that the accuracy of single object grabbing is 96.7%,and the accuracy of multi-target grabbing 94.9%,the algorithm is feasible.