Research On Transmission Tower Key Point Detection Method Based On Deep Learning

In recent years,with the continuous expansion of power grid scale,the traditional manual inspection method can not meet the demand of power inspection,so more and more companies use unmanned aerial vehicle instead of manual inspection,mainly to achieve the detection of insulators and shock hammer and other power components.In this kind of detection,target detection is used,and transmission tower detection is seldom used.In order to better describe the transmission poles and towers in detail,this paper uses the key point detection algorithm,and through the physical relationship between the key points,clearly describes the current state of the transmission poles and towers.This paper focuses on the research of transmission line key point detection,mainly from the following aspects.Firstly,this paper defines a marking method of key points for transmission towers,and makes a data set.In order to set up more reasonable key points of transmission towers,the marking method of key points in human body posture recognition and gesture recognition data set is adopted to calibrate the key points.In view of the small number of data sets,data enhancement technology is adopted to expand the scale of data sets.According to the structure characteristics of transmission poles and towers,Mask RCNN and UniPose human key points detection algorithms are used as the basic network framework of transmission poles and towers.Experiments show that Mask RCNN and UniPose algorithm can achieve 96.3% and 94.8%average PCK in the detection of transmission tower key points.Then,this paper improves the Mask RCNN network to improve the accuracy of key point detection.The first step is to improve the residual network by adding the SE attention mechanism,so that the residual network can pay attention to the feature graph that is more useful for the current task through learning,so as to improve the feature extraction ability of the residual network.FPN,the second step,in view of the characteristics of the pyramid in the fusion of low-level features semantics and high-level semantic features of semantic leakage problem,by adding additional feedback connection,makes the characteristic diagram again after a multi-scale fusion,which can enhance the FPN network,in the form of recursive can effectively enhance the multi-scale characteristics of the power of expression.The third step is to use DIo U LOSS to optimize the regression of the boundary box,which introduces the punishment mechanism,which can effectively improve the regression accuracy and speed of the boundary box,and mark the target area more accurately.Experiments show that the accuracy of transmission tower key point detection can be effectively improved by the above improved method.Finally,the state detection network of transmission poles and towers is built,and the key points predicted by deep learning network and the physical information between them are used to judge the current state of transmission poles and towers.The experiment shows that the transmission pole tower status detection network can effectively judge the normal state,tilt and fall of transmission pole tower.