Research And Experiment On Robot Sorting System Based On Rotating Object Detection Network

In modern warehousing and logistics,it is necessary to keep,store,distribute,and load and unload goods frequently.Therefore,sorting goods is an essential step in logistics.However,with the increase in labor costs,robots gradually replacing labor has become the mainstream trend of future warehousing logistics.Using visual perception technology to endow sorting robots with the ability to randomly sort goods is the focus of intelligent logistics research.At present,most visual robot grabbing systems can only grab commodity boxes that are separated from each other,and the grabbing of commodity boxes that are closely stacked or blocked needs further research.Under complex working conditions,it is difficult for the detection system to obtain position and orientation information such as target plane coordinates,rotation angle,etc.,resulting in poor accuracy of sorting,and easy to miss and mis-select.To address the above issues,this article uses YOLOv5’s rotation box positioning target detection algorithm for predicting the rotation angle of sorting boxes.This algorithm transforms the angle regression problem of the target object into a classification problem,and introduces a circular smooth label angle classification method to achieve accurate prediction of the sorting box body angle;The LCIo U regression box loss function is used to measure the regression loss of the rotating box,which solves the problems of slow convergence speed and poor stability of the rotating target detection model;Then,based on the detection results of the rotating object detection network,the category,position,and rotation angle of the object are obtained.Then the pixel coordinate value and depth value are converted into the robot base coordinate system through the hand-eye calibration method of Intel Realsense camera.Finally,the performance of the improved rotating target detection network and robot sorting system is tested on the experimental platform built in this paper.The main work of this paper is as follows:(1)In order to realize the independent sorting of commodity boxes in the random stacking scenario,this paper builds a robot commodity box sorting system that integrates target detection,rotation angle position recognition and target positioning,and configures the hardware of the entire sorting system,and finally sets the communication mode between the hardware.(2)The calibration model of the robot vision system is constructed.The internal parameters of the Intel Realsense sensing camera are calculated by using the camera calibration method and the homogeneous transformation matrix of the camera coordinate system in the robot base coordinate system is obtained by using the handeye calibration method.(3)When the target object has a large tilt in the image,the traditional horizontal frame positioning cannot complete the precision positioning,and pixel suppression is easy to occur when the object is densely stacked,resulting in missed detection.Based on YOLOv5,this paper introduces the calculation method of angle prediction branch and rotation box regression loss.Through collecting a large number of data sets,we train a rotation target detection algorithm with better model evaluation performance.Finally,based on the improved rotation target detection algorithm,we get the object category,position and rotation angle.(4)In this paper,the robot commodity box sorting experimental platform is built.After the model detection effect is verified,the angle prediction branch is added,and the regression box closely fits the sorting object,and the sorting object characteristics and background are well separated.And the sorting results of several groups of experiments verify that the algorithm in this paper has a higher sorting success rate and more accurate recognition and positioning accuracy than the original YOLOv5 model.