Research On Multi-Target Recognition And Location Method Of Sorting Robot In Warehouse Environment

In order to improve the robustness of robot sorting,the key issue is how to effectively extract important feature information of the target objects,which have diverse,similar,and complex storage features in warehouse environment.Therefore,aiming at the problems of interference within shelves and variety of goods in warehouse environment,this paper uses the deep learning detection method and binocular stereo positioning technology,combined with different grasping strategies of mechanical arm,to carry out the research on grasping of various goods by sorting robot.The main research contents are as follows:(1)Multi-objective dataset building and enhancementThe existing dataset is lacking of openness to various targets in warehouse environment.20 types of target goods are determined through simulating the non-structural features of objects.By placing objects randomly,changing the light intensity,shooting angle and distance,plenty of images are collected to build a multi-target sample library.Then,the images are processed and expanded by translation,random clipping and other methods.(2)Research on target recognition method based on the SSD algorithmThe color and structure of goods are not consistent in warehouse environment by analyzing variety goods.Based on the regression SSD target detection method,4520 samples of target scene annotation are trained by selecting default preselection box and adjusting hyper-parameter method.The two-dimensional image points detected by the target are used as pre-capture points for picking goods,and the recognition accuracy of various goods in warehouse environment reaches over 75%,and the recognition rate reaches 16.8fps,which meets the real-time requirements for visual recognition of sorting-robots.(3)Multi-object grasping point positioning based on binocular stereo visionAccording to the workspace of the manipulator,Eye-to-Hand vision system architecture is adopted,and the double target and hand eye calibration are carried out with MATLAB tools to obtain camera parameters and coordinate system transformation matrix,so as to solve the problem of 3D position information reconstruction of multi-target objects.A binocular epipolar constraint based on region matching is proposed to improve the matching efficiency.In order to improve the fetching efficiency,the principles of selecting fetching points for uneven quality items are optimized.Experimental results show that the accuracy of 3d coordinates of binocular positioning system is 84%.The binocular matching success rate was 89.7%when the target goods were placed more than 4.5cm apart.By adding offsets to the identification points of goods with uneven quality,the success rate of sorting and fetching was increased from 64%to 85%.(4)Research on sorting robot system for warehouse environmentThe hardware,software,structure layout and machine function of the sorting robot was designed after analyzing the tasks of unmanned warehouse application scenarios.According to features of the goods surface,the sorting motion trajectory of a 6-DOF manipulator in the warehousing scene is planned and two end-grabbing strategies of top-face and side-face suck are designed.The simulation and experimental results show that the gripping track of the manipulator is fast and smooth.The average operating efficiency of the arm is 22s/piece,and the success rate of multi-objective sorting is 89%,which realizes the robot system to grasp the target autonomously.The whole working process of the sorting robot is designed.In case of receiving orders and picking goods without any manual intervention,the robot system is required to traverse the shelves and search for the target,and transfer to the turnover box after all the goods are successfully captured.Experiments indicate that the sorting robot can process an average of 20 orders per hour and grasp up to 125 pieces.