Bird Nest Detection Of Catenary Support Based On Deep Learning

With the continuous development of electrified railway,new requirements are put forward for the safe operation of railway.It is an important means to ensure the safe operation of railway to find potential safety hazards in advance and solve them in time through catenary inspection.Bird ’s nest detection is one of the tasks of catenary inspection.Bird ’s nest is usually composed of straw,branch,wire and other materials,which can easily lead to catenary tripping,insulator failure and other accidents.Because it is mainly distributed in the bracket or hard beam,it is difficult to use contact catenary inspection technology to detect it.At present,it is mainly detected by manual inspection,which is inefficient.Using computer image processing technology to automatically identify the nest can greatly improve the detection efficiency.At present,the target detection method based on deep learning has been widely used,and good results can be achieved by training a large number of samples.However,the amount of bird nest data only accounts for a very small part of the catenary image,which belongs to a typical unbalanced data sample,and there will be overfitting problems when directly used for training models.In addition,the bird nest target is small,and the model needs to have strong small target detection ability.Therefore,the acquisition of bird nest data and the accurate detection of bird nests are an urgent technical problem to be solved.The main research contents of this paper are as follows :In the video image of catenary inspection,aiming at the problem of large support target background,an image preprocessing method based on multi-image processing method is designed to enhance the support target body.The binarization algorithm and histogram equalization algorithm are used to enhance the image respectively in the preprocessing,which improves the contrast between the support body and the background.Faster RCNN and Yolov5 are used to detect the bracket,which proves that the pretreatment method improves the recall rate of the bracket,and the detection performance and effect are compared.In order to solve the problem of sparse bird nest samples,this paper proposes a simulation method for bird nest of catenary bracket,which expands a large number of bird nest data and solves the problem of sparse bird nest samples to some extent.Firstly,this paper further locates the bracket components based on the positioning bracket,and compares the effect of linear detection algorithm and deep learning algorithm in the positioning of bracket components.The gamma algorithm is used to adjust the light intensity of the bird nest image,and the bird nest image and the catenary image are fused on different parts of the bracket to obtain the simulation sample.In this paper,thousands of bird-nest data sets are screened out through simulation.Aiming at the small target of the bird nest,this paper proposes a double detector detection method to improve the recall rate of the bird nest.Firstly,the Yolov5 model is used as a detector.The multi-scale mechanism of the Yolov5 model effectively increases its ability to detect small targets,which is suitable for detecting small targets such as nests.The simulation data set is used to train two detectors to detect the bird nest.The first detector detects the bracket and the bird nest in the catenary image,and the second detector detects the bird nest in the bracket image on the basis of the first detector,which improves the recall rate of the bird nest.In view of the discontinuous false detection of multiple frames in the bird nest video,this paper improves a target tracking technology(SORT)to track the bird nest,which effectively reduces the false detection rate of the bird nest.The template matching method is used to accurately locate the bird nest missed by the detector.Finally,the image frames belonging to different bird nests are recorded separately according to the tracking results.