A Study On Workpiece Recognition And Posture Localization Based On Deep Learning

Automatic workpiece grabbing on production line is important for improving the production efficiency in manufacturing industry.However,due to their irregular shapes,uncertain positions,and various posture changes,traditional edge detection,feature extraction and other methods are difficult to accurately identify and locate complex workpieces.In this paper,the workpiece grabbing method based on deep learning is proposed from two directions.A object detection algorithm is applied to the workpiece,and an angle regression network is proposed to realize the 3D pose determination of the workpiece,and the classification,position and angle information are obtained.The other method is based on the single-objective three-dimensional reconstruction method,and the three-dimensional model of the workpiece is recovered according to the two-dimensional RGB image,thereby obtaining the spatial position and posture information,and completing the identification and positioning work of the workpiece.Experiments show that these two methods can solve the problem that the 3D pose of the workpiece cannot be determined compared to the traditional workpiece feature extraction method.At the same time,the method of deep learning is more scalable and the object detection speed is faster.In this paper,the object detection technology and the 3D reconstruction technology are used to obtain the workpiece position and angle information respectively.The effects of the two technologies corresponding to the model in the information acquisition of the workpiece are introduced.The main contents of the paper are as follows:(1)The composition of the experimental platform and the calibration of the camera.It mainly introduces the overall hardware composition of the experimental platform,including the selection of main hardware such as mechanical arm,visual experimental frame,camera,lens and light source.The establishment of the camera model and coordinate system is also introduced.Finally,introduce the principle and steps of camera calibration,complete the calibration of the camera,and obtain the internal and external parameters of the camera.(2)Workpiece recognition and positioning based on object detection algorithm.The main purpose is to realize the migration of the object detection algorithm to the specific workpiece detection.For the problem that the algorithm can not determine the 3D pose in thisapplication scenario,an angle regression module is proposed.It mainly introduces the basic principle of the algorithm and the overall process,network structure,loss function,data set production process and training and testing methods and results.Finally,the other improvements of the algorithm are introduced.(3)Workpiece positioning and posture recognition based on single-objective 3D reconstruction.It mainly compares the advantages and disadvantages of three reconstruction forms of point cloud,grid and voxel.Through the reasonable improvement of the data set,not only the three-dimensional reconstruction of the workpiece with voxel is realized,but also the specific posture can be restored.