Research On Object Localization And Pose Estimation Of Industrial Robot Based On Deep Learning

In recent years,deep learning has shown a strong vitality in the field of computer vision and has great potential in practical applications,but it is less used in the industrial field.In this paper,the problems of object recognition,location and pose estimation of industrial robot based on deep learning are studied.The purpose is to combine the vision of robot with deep learning to extend the object detection task type so that the industrial robot can perform further object grasp or suction operations.First of all,in terms of the capture scene of industrial robot,this paper proposes an location detection model of planar object based on Faster R-CNN.The outputs used in the object location task is converted into 2-points grasp detection.Secondly,in terms of the industrial robot application scenario with suction cups,a plane object location and pose estimation model is proposed based on Faster R-CNN,and the network output layer for estimating the pose is added so that the network can complete the target at the same time.Finally,the robot vision platform is built and monocular two-dimensional vision detection experiment is carried out.In the experimental stage,the performance of the two models is tested,taking Cornell items,industrial PCB and EPFL automobile data sets as examples.In the experiment of detecting the grasp position,the first two data sets are used to detect the 2-points grasps respectively.In the pose estimation experiment,the position and azimuth estimation models are trained according to the accuracy of 10,30 and 45 degrees respectively,and the range of deflection angle of PCB in 2D plane and the azimuth range of EPFL vehicle are obtained.The experimental results show that the recognition accuracy can reach 98%,the average coincidence degree of the grasp location range and the standard true value can reach 73%in the grasp position detection task.In position and attitude estimation experiment(deflection/azimuth detection)the accuracy of the average position estimation can reach 97.5%/90.6%,93.6%/88.2%and 89.7%/82.6%respectively,and the recognition accuracy can be kept above 98%.The results show that the two models proposed in this paper have practical value in the task of position and pose estimation and position detection of target in 2D plane.