Research On Transmission Line Foreign Objects Intrusion Recognition And Localization Based On Deep Learning And Binocular Vision

The intrusion of foreign objects in the transmission line channel can lead to short circuit,flashover or tripping of the line,which may affect the reliability of power supply.Foreign objects are usually found by operation and maintenance personnel screening inspection images,which is inefficient and unable to obtain spatial information of foreign objects.In this thesis,Based on the methods of deep learning and binocular vision,the captured transmission line channel images were analyzed to identify and locate the foreign objects.The main work is as follows:The data set of intrusion foreign objects in transmission line images was built,the training strategies of objects detection network and semantic segmentation network were optimized,and the model of foreign objects detection and segmentation were obtained.By replacing the backbone,introducing the attention mechanism and conditionally convolution module,the Retinanet objects detection network was improved.The classification accuracy of mF-Score on the test set has achieved 0.884,and the ablation experiment was used to verify the effectiveness of the improvement.At the same time,the Deeplabv3 + semantic segmentation network was used to realize the segmentation of foreign objects in the transmission line image.mIoU on the test set is 0.82,and the feasibility of using rectangular box annotation instead of mask annotation for network training was studied.The results show that when the proportion of mask annotation is only 50 %,the mIoU of Deeplabv3 + model can still reach 0.71.Through binocular camera calibration,the internal and external parameter matrix of camera were obtained,the relationship between camera imaging coordinates was established and the binocular image correction was completed.Based on AD-Census stereo matching algorithm,an adaptive weight MAD-Census stereo matching algorithm was proposed by using cost function of larger receptive field and introducing adaptive weight to characterize texture region features.The average mismatch rate of the proposed algorithm on Middlebury dataset is reduced by 0.81% compared with AD-Census algorithm,furthermore,the stereo matching results in the edge regions,the repeated texture regions and the weak texture regions of simulated transmission line scene is better than that of AD-Census.The disparity map obtained by the improved algorithm can reflect the spatial information of transmission line scenes and foreign objects preferably.Combining with deep learning recognition results and Euclidean point cloud clustering algorithm,3D Point Cloud construction based on binocular vision was adopted to extract foreign objects point cloud and remove background point cloud.The point cloud model of foreign body was established to obtain the 3D information such as the center coordinates and the maximum size of foreign objects.An example of transmission line foreign objects intrusion recognition and location was analyzed,and the indices of deep learning recognition were obtained;The error analysis of the location results was carried out,it shows that the spatial measurement error of foreign objects point cloud is basically less than 5% within the depth measurement range of 15 m,which verified the feasibility of the algorithm in this thesis.