Research On Deep Learning Based Workpiece Unordered Sorting System

In recent years,the industrial sector has witnessed significant growth in the use of robotic arm gripping,owing to the rapid advancement of industrial intelligence.As the first step in the gripping operation,the generated gripping pose guides the subsequent path planning and gripping execution,and its accuracy determines the probability of successful gripping.In unstructured environments,where multiple targets are stacked and workpieces are placed randomly,it is difficult for the robotic arm to complete the gripping task accurately and stably.In order to improve the intelligence of the grasping system and realise the unstructured sorting of workpieces,this paper proposes a two-stage grasping detection algorithm based on deep learning.The algorithm optimizes the target detection and grasping detection network in steps to achieve the detection of obscured targets in multi-target scenarios.The main research work is as follows:First,the overall framework of the system is designed and the hardware selection of the grasping system is completed according to the experimental environment and the requirements of the grasping task.The camera calibration was completed by the Zhang Zhengyou calibration method,and the hand-eye calibration was completed by the eye-out-of-hand grasping method to realise the conversion of the grasping information under the 2D image coordinate system to the grasping poses under the 3D robotic arm space.Secondly,the model structure of the YOLOv5 target detection algorithm is investigated.To address the problem of low detection accuracy due to occlusion between workpieces in complex environments,the YOLOv5 model is optimally designed by thinking from four perspectives: data enhancement,network structure,loss function and non-maximal value suppression algorithm.The Grid Mask algorithm is used for data enhancement of the training images to improve the robustness of the algorithm;ASPP and Bi FPN modules are introduced to improve the feature extraction capability and CBAM attention mechanism is added to improve the feature utilisation;the deficiencies of the loss function in the detection of occluded targets are analysed and the Repulsion Loss rejection factor is introduced to improve the regression accuracy of the bounding box;Adding a density function to Soft-NMS to dynamically adjust the threshold and reduce the probability of false detections and misses.Experiments show that the detection accuracy in the VOC dataset is improved by 2.1%compared with the original algorithm,and the detection accuracy in the homemade workpiece dataset reaches 94.2%,effectively improving the detection capability of obscured targets.Then,the pixel point-based GRCNN grasping detection algorithm is investigated.A twostage grasp detection algorithm is designed for the problems of unknown target types,severe environmental interference and repeated detection for grasping under multi-target occlusion.The type of target and the grasping area are first obtained through the target detection network,and then the optimal grasping frame of the target is obtained through the grasping detection algorithm,so as to achieve the grasping of the specified target.At the same time,the network structure of GRCNN is improved by introducing CSP and ASPP modules to improve the performance of grasping and detecting multi-scale target objects.And the a priori knowledge of target detection is used in post-processing to filter out non-conforming grasp prediction points to improve the accuracy of grasp detection.Experiments are conducted on a home-made occluded artefact dataset,and the results show that the grasping accuracy reaches 96.2%,satisfying the grasping detection in a multi-target occlusion environment.Finally,a vision-based robotic arm grasping physical platform was built and set up for single-target grasping experiments and multi-target grasping experiments.The results showed that the grasping success rate was 91.25% in single-target scenarios;90% in scattered placed multi-target scenarios;and 82.25% in densely occluded multi-target scenarios.The system has been verified to be practically feasible.