Research On The Key Technologies Of Rail Traffic Safety Video Detection Based On Deep Learning

With the development of China’s economy and the expansion of cities,the importance of rail transit in urban transportation system has become increasingly prominent.However,as various safety problems have also emerged,the application of video image processing technology to rail transit safety system is essential.This involves many cutting-edge subject technologies,especially the application of deep learning technology in the object recognition field of image processing,which solves the problems of poor performance and poor robustness of traditional object detection algorithms.The methods to detect dangerous phenomena in rail traffic system are mainly studied in this paper,and using the object detection algorithms which are based on deep learning to realize the determination procedure of dangerous phenomena in rail transit scene and complete the system framework.In this paper,the object detection algorithms which is based on image processing are studied.The time averaging method is used to establish the background model,the background difference method is used to obtain the foreground targets,and the use of image morphology processing algorithms enable foreground targets to become more complete.At the same time,the concept of dispersion is applied to the object detection of image processing to improve the detection accuracy.The methods to detect the dangerous phenomenon of pantograph fault arc-drawing in rail transit scenarios are studied in this paper.On the basis of gray-scale image characteristics and pulse time characteristics of pantograph fault arc-drawing,the arc-flashing target detection method based on state tracking is given.On this basis,combined with the dispersion characteristics of pantograph fault arc-drawing,the arc-flashing target detection method based on dispersion analysis with better performance is given,which is suitable for more complicated rail transit scenarios.The methods to detect the dangerous phenomenon of foreign body intrusion and cross-border based on YOLOv3 are studied in this paper.Several classic object detection algorithms based on deep learning are analyzed comprehensively,among which YOLOv3 is selected as the network of object detection in this paper because of its high accuracy and high speed.The parameter training method for specific objects in the YOLOv3 network is given,and the training parameters are used to complete the detection and classification of objects in the rail traffic scene.The integral rail transit safety video detection system is designed and implemented in this paper.The hardware framework of the system is realized,including video data acquisition module,server image processing module and client alarm receiving module,which provides data support for the acquisition of high-definition video data and subsequent video image processing.The software framework of the system is realized,including dome camera drive module,monitoring pre-processing module,dangerous phenomenon detection module and alarm messages transmission module.The construction of the hardware and the programming and test optimization of the software framework are completed.The test of capturing the evidence for the dangerous phenomenon in the actual scene is completed and the result is good.