Design And Implementation Of A Stacked Workpiece Recognition And Positioning System Based On Deep Learning

In industrial production,the workpieces produced by casting and other processes are generally stacked in hoppers or pallets,and then sorted for subsequent finishing.Conventional manual sorting methods have some disadvantages,such as poor working environment,high labor intensity,sorting errors caused by fatigue of repeated work,etc.In the automatic production line,the specially designed sorting equipment often has some shortcomings,such as loud noise,high failure rate and poor adaptability.In order to improve the sorting efficiency,improve the sorting quality,reduce the labor intensity of workers and improve the working environment,as well as meet the needs of small batch,multi-variety and mixed production,this paper studies and develops a set of stacked workpiece recognition and positioning system based on deep learning for the field of robot sorting.The main work is as follows:(1)The principle of Faster R-CNN model,which is commonly used in the field of target detection,is studied,and the stacked workpiece detection experiment is carried out.The stacked workpieces are divided into severe,moderate and light scenes.Experimental results show that the Faster R-CNN model has large fallout ratio detection and false detection rate in both moderate and severe scenes,and the detection effect is poor.(2)Improve Faster R-CNN.ResNet and SENet are combined into SE-ResNet to construct a feature extraction network,which improves the problem of insufficient feature information extraction of workpieces in stacked state.Based on Soft-NMS algorithm,Bin-Soft-NMS algorithm is proposed,which improves the problems of low model recall rate and suppression of boundary box score drop of adjacent stacked workpieces caused by inaccurate threshold by dynamic threshold.The loss function is constructed by Reference Loss,which solves the problem that the prediction box of workpiece is affected by the real box and prediction box of other stacked workpieces.Experimental results show that the average detection accuracy of the improved model is 86.6%in stacked workpiece scene,and it is improved by 1.4%,3.5%and 5.125%in light,moderate and severe scenes respectively,with an average improvement of 3.34%.(3)The architecture of stacked workpiece recognition and positioning system is designed and implemented,and the calibration of KinectV2 camera is completed.The hardware equipment of the system is selected,and the software architecture and main functional modules are designed in detail.The system adopts an improved Faster RCNN deep learning network model to improve the workpiece detection accuracy.JETSON NANO development board is adopted to reduce cost.KinectV2 camera is used to solve the problem of lack of depth image data in traditional plane target detection.The overall flow of unordered sorting process of stacked workpieces based on ZrARM robot is formulated and realized.The calibration principle of KinectV2 camera is studied,and the color camera and depth camera of KinectV2 camera are calibrated by Zhang Dingyou’s calibration method,so as to obtain the internal and external parameter matrices and distortion parameters of each camera.(4)System integration and hybrid object grasping experiments are carried out.The experiment of unordered sorting of stacked workpieces based on ZrARM robot is completed.The experimental results show that the picking success rate of the system reaches 87%,which can basically meet the needs of industrial production under certain conditions.