Vision Based Localization And Perception Of Inspection Robot

Electric power is an important cornerstone of social operation,and the stable operation of power plants is the guarantee of power supply.The daily operation of power plants requires manual scheduled inspections.This inspection method has the problems of high labor intensity,unstable quality,and poor working environment.The introduction of inspection robots can reduce the labor investment and improve the quality and effectiveness.The power plant has a large number of equipment and pipelines.These equipments need to be visually observed to see whether there is a fault.However,traditional manual inspections lack automatic processing of inspection image data,and fail to effectively detect faults that require historical archive comparison.The convolution neural network algorithm based on deep learning can classify the huge picture data quickly and accurately,and improve the quality of patrol inspection.In order to make the inspection more efficient,this paper studies the autonomous positioning and navigation of inspection robot based on monocular camera and visual perception based on deep learning.The perception part includes target detection and classification.Robot system and target perception are in the same system environment,and the algorithm can be transplanted well.The research contents and achievements of this paper are as follows:(1)This paper summarizes and analyzes the methods and shortcomings of current inspection robot navigation and positioning,and proposes v-slam method using monocular camera for position and orientation estimation and navigation of inspection robot.Use the ORB-SLAM algorithm to extract ORB feature points through the monocular camera,build a sparse 3D feature map,and complete positioning,composition,path planning,navigation in a laboratory environment.(2)Aiming at insufficient computing performance of mobile robots,insufficient real-time detection of target detection algorithms,and lack of targeted object detection in power plant environments,K-means clustering algorithm is used to cluster power plant data to determine the number and size of anchor boxes.An improved YOLO target detection algorithm is proposed.Compared with the original algorithm,the accuracy rate is improved,the detection speed is increased by 22%,and the training time is reduced by 26%,which proves the effectiveness of the algorithm.(3)Because the features of the target detection algorithm are not deformed and robust,similar devices that perform different functions at different locations will be detected as the same object.In order to distinguish this,first set the standard direction of each object,filter the data of similar angles according to the target detection threshold,select the supervised learning argmax classification method,extract the background region features through ResNet50 network,and Softmax regression classify the object.The training speed and the accuracy of the model are improved by using data standard normalization,learning rate warmup,and early stopping.The accuracy of the Validation set is improved by about 5%.