Research And Realization Of Bird’s Nest Detection Method For High-speed Railway

By the end of 2020,the total mileage of our country’s railway operations has reached146,000 kilometers.In the spring and autumn seasons,a large number of birds build nests on the high-speed rail catenary.The voltage of the high-speed rail catenary is about 27.5k V.When the branches of the bird’s nest are damp,the catenary is often short-circuited,which brings risks to train operation.At present,the main method of detecting the bird’s nest in the railway power supply section is to manually view the inspection videos recorded by the C2 camera,but this consumes a lot of manpower,also has low recognition efficiency.The difficulty of bird’s nest detection lies in the small size of these branches,fewer pixels in the image,fewer features,and confusion with lines,which are difficult to identify.This paper has designed a bird’s nest detection method based on deep learning for the inspection videos.First of all,this paper analyzes the video frames of train inspection provided by railway department,and detailed researches were carried out on the problems of complicated background,fog,haze,insufficient light and the tilt of catenary pillars in the image.The problematic images which were collected were performed pretreatment of defogging and tilt correction.Secondly,identify the catenary and pillar areas.This paper uses the improved YOLO-v4 algorithm to identify and extract the catenary and pillar areas.On the basis of re-clustering the anchor frames,the features information of small objects in the deep neural network is enhanced,and the recognition accuracy of small objects is improved,and the recognition time is optimized.Third,an improved StyleGANs network is designed to recognize the bird’s nest.On the basis of generating high-resolution images in different backgrounds,a classifier is added to the discriminator to extract detailed information from the generated image and the real image.The characteristics of the samples are classified to learn whether there is a bird’s nest,and the quality and diversity of the generated images are evaluated at the same time.The classification effect of the discriminator model is tested on the test set.According to the results,the recall rate is used to evaluate the quality of the model.The recall rate of the algorithm in this paper reaches 96.5%.Finally,this paper develops and implements a nest detection system,which integrates the algorithms of this paper and combines it with the identification of pillar numbers.It realizes the functions of videos analysis and frame acquisition,preprocessing of catenary images,bird’s nest detection area extraction and bird’s nest recognition.