| With the rapid development of the economy,the scale of the manufacturing industry is also constantly increasing.As a key factor in the improvement of the manufacturing industry,industrial robots are widely used in various industries of the manufacturing industry due to their advantages of saving human resources,reducing production costs,and improving production efficiency.Improving the intelligence level of industrial robots will help promote the transformation of traditional manufacturing industry to intelligence,automation and Digital transformation,further enhance the international competitiveness of "Made in China",and then achieve the strategic goal of becoming a powerful manufacturing country.The robotic arm grasping work is the fundamental operation of industrial robots in production processes such as stacking,handling,and sorting.By changing the traditional grasping method and combining visual technology with industrial robots,robots can accurately recognize the category and pose of target objects in complex environments,improving their intelligence level.This article studies the target recognition and grasping prediction technology of visual industrial robots,and proposes a recognition and grasping prediction method based on deep learning technology.The main content of the research is as follows:(1)In response to the problem of collecting visual image information for unknown irregular objects,a binocular vision system was established and the camera was calibrated using the Zhang’s checkerboard calibration method.The transformation matrix between internal and external parameters and coordinate systems was obtained,and an eye in hand grasping platform was established to complete the robot human hand eye calibration experiment.(2)In response to the problem of object detection algorithms in the experimental environment,a D-YOLOv4 object detection network model is proposed.This model is mainly based on the YOLOv4 network structure,and the standard convolutional layer in the network is replaced with a deep separable convolutional layer to improve the network running speed.At the same time,the size of the prior box when predicting the boundary box is improved to improve algorithm performance.Then,the object detection dataset created in this article is utilized,after training the improved target detection network model and comparing it with the original algorithm,it was found that the detection speed of the improved network model increased by 4.6 fps,and the average detection accuracy increased by 1.4 percentage points.(3)Aiming at the problem of grasping pose detection of unknown irregular targets,a grasping detection network model D-YOLOv4-grasp based on D-CSPMarknet-53 is proposed.This model uses the five dimensional parameter representation method to facilitate the output of grasping coordinates and angles,and uses regression ideas to determine the loss function of the network model,which is trained by grasping data sets in CGD and Jacquard,The experimental verification and analysis using target detection dataset images show that this method can achieve effective object grasping prediction.(4)In response to the experimental verification of the algorithm,an overall scheme of the visual industrial robot recognition and grasping system was designed.A robot recognition and grasping experimental platform was built using hardware such as FANUC industrial robots,computers,binocular cameras,pneumatic grippers,controllers,and teaching aids.At the same time,the robot communication module design and robot grasping program programming were completed in terms of software,On the established experimental platform,experiments were conducted on the selected grasping objects.The experimental results showed that the total success rate of the robot recognition and grasping system was 82.5%,verifying the effectiveness of the object detection algorithm and grasping prediction network proposed in this thesis in the human recognition and grasping experiments of visual industrial robots. |
PDF Full Text Request