| Modern manufacturing industry has put forward higher requirements for the intelligence and flexibility of production process.Automatic assembly and multi-workpiece parallel processing and other scenarios have a wide range of requirements for multi-type workpiece the recognition and pose estimation technology.However,due to the lack of industrial parts surface texture and other reasons,The results of existing algorithms are often unsatisfactory.Current workpiece recognition algorithms usually require manual defining and extracting features,and poor ability to recognize multi-variable types and low-texture artifacts.In terms of positioning,there are the hardware cost is high,the amount of calculation is large,can not achieve arbitrary 6D pose positioning and other problems.To solve the above problems,this paper proposes a method based on binocular vision,combined with deep learning and contour reconstruction to recognize and estimate the pose of low-texture and multi-type workpiece.The main work and achievements are as follows:(1)A binocular vision system was set up and calibrated.Firstly,the camera coordinate system and the principle of binocular vision are introduced.Setting up the binocular vision system required by the experiment.Then using Matlab to calibrate the internal and external parameters of the binocular camera,and the reprojection error is only within 0.22 pixels.Finally,according to the calibration results,the acquired images are distorted and stereo corrected.The ideal binocular vision system is obtained.(2)Developed the workpiece recognition algorithm based on YOLOV4.The principle and network structure of YOLOv4 algorithm and the improvement points compared with YOLOv3 are studied.The network model is set up by Python and Pytorch under Windows10 system.Collectting and labeling the workpiece data set.Then YOLOv4 algorithm is optimized by priori box.Training YOLOv4 network by using transfer learning method.Finally,the m AP index obtained by the experimental test reached 84.9%.(3)A 6D pose estimation algorithm based on workpiece contour point cloud reconstruction is proposed.Firstly,line segments were extracted from the region where the workpiece was located,and a multi-constraint line segment matching algorithm combining with the results of YOLO target detection was proposed to match the set of left and right line segments,reconstruct the spatial equation of line segments and break the contour point cloud of the workpiece.Finally,6D pose was obtained by registration with the template point cloud.The final experimental results show that the position error is less than 1mm and the attitude error is less than 0.8°.(4)Multi-workpiece recognition and pose estimation system construction.In terms of hardware,an image acquisition and processing device is set up.In terms of software,using C++ programming language and Qt interface framework,Hikvision SDK,Open CV,PCL library,developed the corresponding software.Finally,a comprehensive experiment is designed to verify the effectiveness and superiority of the proposed method. |
