Research On Nuclear Waste Sorting Method Based On Machine Vision

With the implementation of the "13th Five-Year Plan" of "safe and efficient development of nuclear power",China’s nuclear power operation and installed capacity under construction will reach nearly 88 million kilowatts by 2020,and the safe and efficient disposal of nuclear waste is the key to ensuring the sustainable development of nuclear energy and the safe application of nuclear technology.This topic around the nuclear waste disposal in the back and sorting operation needs,in view of the current nuclear waste sorting teleoperation robot,the only existing remote control operation by monitoring the camera by the operating personnel of information perception ability is insufficient,low efficiency of operation,the research of nuclear waste sorting method based on robot vision,the robot back to fetch the introduction of nuclear waste and the sorting,using the target detection network optimal grasping pose estimation and cascade of sorting algorithm.A cold experiment platform and nuclear waste simulation data set were built to verify the effectiveness of the algorithm.The main research work of this paper is as follows:(1)The framework of the robot sorting system was constructed,and the hand-eye calibration model was established through the Universal Robots5 manipulator and the Inter Realsense stereo camera to complete the hand-eye calibration of the system.(2)A nuclear waste detection and classification model based on the improved Retinanet network is designed.By improving the structure,the number of feature extraction layers of Retinanet network is reduced and multi-scale receptive field modules are added to improve the network speed and accuracy.Training and testing were carried out on PASCAL VOC and nuclear waste simulation datasets respectively.On PASCAL VOC datasets,the mean mean accuracy MAP =79.5% and the frames per second FPS=31.95.On the nuclear waste simulation data set,the RGB and RGD diagrams were input for training and testing,and the RGD diagram performed better,with the accuracy and speed of 95.14% and 37.96,respectively.(3)Estimation of the optimal grasping position and pose of nuclear waste.In view of the unpredictability of nuclear waste and the difference of appearance and material,an optimal capture pose estimation network model based on RESNET-50 backbone network and FPN multi-scale features was designed.With RGD information as input,a capture candidate box was generated.The coordinate of grasping direction is mapped to the classification task of grasping direction,and the optimal grasping pose of the target is obtained.The grasping pose estimation experiment is carried out based on the Cornell grasping data set,and the accuracy of the grasping pose estimation reaches96.9%,which verifies the effectiveness of the model.(4)Experiments and analysis of robot sorting system were carried out.The cascade of target detection network and pose estimation network is completed,and the function and performance of the algorithm are tested and verified on the physical experiment platform built.