| Traditional garbage cleaning in scenic spots mainly relies on manual work,which has low efficiency,high cost and some garbage is harmful to human body.With the rapid development of machine vision and robot technology,the use of robots instead of manual garbage sorting has become a development trend,so the study of scenic spot cleaning robot based on machine vision has important significance for the unmanned and intelligent garbage cleaning in scenic spots.In this paper,ROS operating platform and deep learning method were used to build a visual system with garbage classification and recognition function,and the system was deployed to the mobile robot experimental platform.Through hand-eye calibration and Move It! To realize the garbage sorting and recycling task in scenic spots,the specific contents are as follows:(1)In view of the problems of complex structure and low accuracy in most of the current garbage image classification networks,this paper uses lightweight model Mobile Net v3 as the backbone network of garbage target detection,and improves the garbage classification network through classification layer structure optimization,data enhancement,RAdam optimization and transfer learning.The improved Mobile Net v3 network has achieved an average accuracy of 99.6% on Trash Net,an international open data set,and its performance is better than other garbage sorting algorithms proposed in the literature.(2)Considering that there are few public garbage data sets at present,and most of the existing data sets have few types and uneven distribution,this paper made a garbage data set(4 categories and 25 categories)based on the data set(4 categories and 40subcategories)provided by Huawei Cloud garbage classification Competition and combined with the garbage background of scenic spots.In this paper,YOLO v7 is used as the basic network for garbage target detection in scenic spots.In view of the complex network structure and large number of parameters of YOLO v7,the ELAN-CSP backbone network in YOLO v7 is replaced by the improved Moblie Net v3.The common convolution of feature fusion network(FPN)is replaced by deep separable convolution.In addition,the addition of SE attention module can enhance the feature expression ability of YOLO v7 network and reduce the impact of precision reduction caused by the reduction of the number of model parameters.The results show that compared with the original model,the parameter number and floating-point operation(FLOPs)of the improved YOLO v7 garbage detection network are reduced by 49.6% and 73.5% respectively,which significantly reduces the network structure,makes the model lighter and the training speed faster,and finally the training model has a m AP value of 90.8 on the self-made data set.(3)Deploy the trained improved detection model to the mobile operating robot platform,complete the classification and recognition of garbage in the scenic spot,and determine the garbage targets and corresponding placement points according to the recognition results.The coordinate transformation matrix of the target and the robot base was obtained by the hand-eye calibration algorithm,and combined with the Move It! The kinematics calculation module solves the grasping position of the gripper target and the Angle of each joint target,and uses RT-Connect algorithm to complete the garbage sorting grasping action.The improved YOLO v7 model proposed in this paper provides a lightweight real-time and efficient detection algorithm for garbage classification in scenic spots.Combining ROS and robot platform,it provides a new idea for unmanned garbage classification in scenic spots. |
