Visual Navigation Method For Electric Power Inspection Robot Based On Image Preprocessing And Semantic Segmentation

The power inspection robot is an important tool for periodic inspection of power equipment,for which a stable and reliable navigation method is the basis to carry out inspection work.Electric power inspection robots realize navigation by collecting navigation path images,which can improve inspection efficiency instead of manual inspection methods.Due to the complex outdoor conditions of electric power sites,which are easily affected by light and bad weather,the traditional image processing method is not efficient for visual navigation of inspection robots.This thesis proposes a visual navigation method of electric inspection robots based on image preprocessing and semantic segmentation.This thesis mainly studies the restoration,enhancement,segmentation,intersection bifurcation sign recognition and offset calculation of the navigation path image for the inspection robot.First,starting from the application scenario of the electric inspection robot,the layout of the navigation path is reasonably planned.The disturbance to the navigation path image caused by the surrounding environment when the robot works in the electric field is analyzed,and an image restoration method based on the combination of constrained least squares filtering and median filtering is proposed to restore motion blurred images.Aiming at the influence of strong light,weak light,uneven illumination and thin snow coverage on the navigation path image,the gamma correction method is improved to achieve image enhancement.A navigation path segmentation method based on the improved FCN-DenseNet network is proposed where the main structure of the DenseNet network is studied.According to the characteristics of the navigation path sample distribution in complex environments,the activation function and loss function are improved.Combined with image preprocessing and improved network,the segmentation accuracy and precision are increased by 5.92%and 5.07%respectively.In order to improve the accuracy of the inspection robot to check the running status and wear of power equipment at fixed points,and to meet the needs of the model to work on mobile embedded devices,the target detection network YOLO v4 was improved,whose feature extraction network was replaced with MobileNet v1.The network uses the K-means++clustering algorithm to optimize the target detection frame to achieve the goal of reducing the weight of the network and improving the recognition accuracy.After a large number of experiments and comparisons,the improved YOLO v4 network target detection frame rate reached 29fps,and the average accuracy and mean mAP increased by 4.62%,which can meet the real-time and accuracy requirements of inspection robot operations.Finally,based on the edge detection method,the incompletely segmented navigation path is restored,and the RANSAC line fitting and the offset calculation method based on the weight of the pixel center point are used to obtain the most realistic navigation path change trend.Image preprocessing,semantic segmentation,target detection and other methods are used to identify navigation paths and intersection signs,and calculate the offset of the navigation path,which realizes real-time adjustment of the robot’s motion posture and improves the intelligence of power inspection robots.