A Deep Learning-based Robotic Trash Sorting Method

In recent years,the problem of urban garbage is becoming more and more serious with the development of urbanization.The classification of garbage has become the focus of the society now.Under the current situation of China,the sorting of garbage by category is mainly done manually,which has the disadvantages of low efficiency,high cost and high labor intensity.With the rapid development and application of robot technology,using robots for garbage sorting has become a development trend.Several foreign companies have launched mature robot garbage sorting systems and solutions.However,the research on robot garbage sorting technology in China is relatively late and is still in its early stage.There are no mature products or solutions yet.It is urgent to research and solve related technologies of automated and intelligent garbage sorting on account of the serious garbage problem.Therefore,this paper focuses on the key technologies of garbage recognition and classification,garbage tracking and robot garbage sorting.The specific work is given as follows:Firstly,in order to solve the problems of low accuracy,poor adaptability and slow detection speed existing in the traditional garbage classification method,the YOLOv3 network model with fast speed,high accuracy and strong generalization ability is selected for garbage classification and recognition.Experiments show that the mean average precision(m Ap)of the YOLOv3 network on the garbage data set collected in this paper reaches 97.4% and the detection time only takes 25 milliseconds.Secondly,a multi-target tracking algorithm combined with YOLOv3 target detection algorithm and kernel correlation filtering algorithm is proposed for the moving garbage targets transported on the conveyor belt.At the same time,a virtual detection line counting method is proposed to realize the statistics of the garbage target.Experiments show that the kernel correlation filtering algorithm has good tracking effect and good fault tolerance for a small amount of occlusion.For the test video,the accuracy of the proposed multi-target tracking algorithm is 86.67%,the precision is 92.17% and the processing speed is 23.6 frames per second.In addition,the counting accuracy of virtual detection line can reach 98.1% in the case of no occlusion and 96.3% in the case of partial occlusion.Thirdly,a visual guidance method for robot grasping is proposed,which can make robot grasp accurately.It is based on YOLOv3 target detection algorithm,kernel correlation filtering algorithm and Kalman filter algorithm.This method can make full use of the advantages of the three algorithms and can adapt to variable speed garbage targets.Finally,the robot garbage sorting system is built and the YOLOv3 network is pruned and transplanted to the Xavier platform.Experiments show that the pruned network can still complete the detection and recognition of garbage targets well in the experimental environment.The accuracy and recall rate of detection are higher than 95%.In addition,the success rate of static garbage sorting is 97.2% and the average time is 4.26 seconds.For moving garbage sorting,the success rate is 92% and the average time is 4.7 seconds.The effectiveness and availability of the garbage sorting system have been verified and the garbage sorting system is expected to be used in real garbage sorting in the future.