Research And Implementation Of Key Technologies Of Collaborative Robot Sorting System Based On Machine Vision

Collaborative robots are robots that can directly interact with people in the work area.They are an excellent platform for intelligent robots and are now widely used in various sorting operations.However,the existing collaborative robot sorting system is still not smart enough to adapt to most factory assembly line sorting operations.There are three main difficulties:the existing factory assembly line sorting operation is mainly dynamic sorting,which requires high real-time and tracking capabilities.Most of the parts on the assembly line are without texture or less texture,which results in the inability to manually design texture features and multiple parts occlude each other and have the similar characteristics,which easily leads to the loss of tracking of the target part;the collaborative robot needs to obtain the object category and the spatial six-degree-of-freedom(6D)pose estimation for the object and the characteristics of non-textured parts lead to low recognition accuracy of existing algorithms and low success rate of capturing target parts;there are a large number of special-shaped parts generated after processing or unrecognizable parts on the industrial assembly line,which requires manual secondary recognition and inputting identification results for the robot to perform secondary sorting.However,the existing interactive system is mainly based on the mouse and keyboard,and the interaction is not direct,which can easily lead to missed inspections,false inspections and lower operating efficiency of the sorting system.In this paper,aiming at the three difficulties of collaborative robot dynamic sorting of industrial parts,the key technology of the entire collaborative robot sorting system has been researched and implemented:1)An algorithm of the estimation of the hand pose and the hand shape based on point cloud and graph network is proposed to replace the traditional human-computer interaction,and improve the operating efficiency of the overall system through hand pose estimation.2)A dataset is established for non-textured parts,and a 6D pose estimation algorithm based on semantic segmentation and point cloud denoising autoencoder is proposed,which can accurately identify the category and pose of static non-textured parts.The algorithm provides pre-acquired category and pose information for tracking target parts later.3)A single target pose tracking algorithm based on Siamese Network and point cloud registration is proposed,which can track the position information and posture information of the target part in real time.The algorithm improves the success rate of the target part grabbing of the entire sorting system.4)A modularized collaborative robot assembly line sorting system is built,and it can be tested and applied in actual scenarios.