Research On Abnormal Event Detection In Railway Scene Based On Surveillance Video

Railway transportation is highly valued for its advantages such as strong capacity and low cost.In recent years,China’s railway construction has made great progress,the operating mileage has increased significantly,and the operating environment of trains has become more complex.Because the periphery of the railway line is not completely closed,the abnormal events which threaten the safe operation of the train caused by human or natural factors often occur.At present,the detection of abnormal events in railway scenes mostly depends on hardware equipment,which is costly and less intelligent.Most of the existing methods for detecting abnormal events in railway scenes based on video surveillance cannot solve the problems of various forms of abnormal events and many interferences such as changes in lighting.At the same time,due to the inherently diverse and harmful characteristics of abnormal events,it is impossible to produce classified data sets,which increases the difficulty of abnormal event detection task.Therefore,it is of great significance to study an intelligent railway scene abnormal event detection method.This paper mainly focuses on the abnormal event detection method under the railway scene,and effectively solves the problem of abnormal event detection and location.The main work and results of this paper are as follows:Firstly,in view of the characteristics of abnormal event detection task in railway scene,we study an abnormal event detection method based on Generative Adversarial Networks which can be applied in various railway scenes(such as platform,railway crossing,etc.).This method is improved on the basis of Generative Adversarial Networks,and only normal event samples are used as the training set,so it is difficult to reconstruct abnormal event samples that do not participate in training.Therefore,this method can realize the detection and location of abnormal events through the difference between the reconstructed image of the target sample and the real image,and effectively solve the problem that it is difficult to obtain the classification data set.Secondly,in order to improve the ability of the network to obtain the interframe information of the input video sequence,the generation network in the above Generative Adversarial Networks is improved.The abnormal event detection framework based on 3D-GAN proposed in this paper takes RGB flow and optical flow as the input of two streams,and uses 3D convolution to replace the ordinary convolution in the image generation network,which effectively improves the ability of the network to obtain the motion information of the target sample.We use deep separable convolution to optimize network structure and reduce network computation.In the experiment of this paper,we have realized the spatial and temporal feature fusion module,which adaptively gives different weights to each feature vector in the double-stream input bottleneck layer,so as to better aggregate and generate the spatial and temporal information in the coded part of the network.Finally,the railway crossing scene was selected as the task objective.Combining with the actual requirements of abnormal event detection in the railway crossing scene,we designed and developed the railway crossing abnormal event detection system.We proposed and realized the railway crossing traffic state detection algorithm to improve the operating efficiency of the software system.We developed the corresponding man-machine interface and completed the task of abnormal event detection in railway crossings.The abnormal event detection network model designed in this paper has been tested in UCSD-Ped2,CUHK Avenue and real railway scene data sets,and good experimental results have been obtained,which verifies the effectiveness of the proposed method.