Detection Of Abnormal Group Behavior In Subway Application Scene

Subway stations are public places with complex scenes,high crowd density and high passenger flow mobility in cities,so the safety problem can not be neglected.How to effectively use the subway video surveillance system to detect the population status in complex scenes in real time and reduce the loss caused by abnormal group behavior,has important practical significance.Unlike traditional motion detection or classification models,subway surveillance video has many characteristics,such as complex scenes,many types of abnormal behaviors but a small number of them.The detection model based on unsupervised learning and fully supervised learning can not meet the practical application requirements.Therefore,group abnormal behavior detection models are built based on the semi-supervised learning theory in this paper,that only needs normal group behavior samples in the training set,and the traditional manual feature selection models are replaced with the end-to-end model of deep learning.The main contents of this paper are as follows:1.At present,most of the mainstream abnormal behavior data sets come from foreign countries,which are different from the actual monitoring scenes of domestic railway stations.After the actual investigation,monitoring videos of wuxi subway sanyang square station are chose as the data source in this paper,and Wuxi Subway dataset is also been built.The dataset has complex scenes,different directions of dense crowds and changeable behavior,which can well reflect the real monitoring scenes of domestic subway stations.2.Based on the properties of convolutional autoencoder which can reconstruct normal sample space and feature compression.In this paper,a group abnormal behavior detection model based on convolutional autoencoder is established.Multiple video frames are selected by sliding window on the time axis,and multiple video frames are superimposed as the input of the convolutional autoencoder.According to the relationship between reconstruction error and threshold value,abnormal behavior events can be determined.Different from the traditional convolutional autoencoder model,the model contains two branches in the decoding stage,which are used to reconstruct the current frame sequence and the past frame sequence respectively.In the training phase,these two related tasks learn the motion trend of normal behavior target by sharing the parameters in the coding phase,which improves the model's generalization ability.3.The convolutional autoencoder model extracts motion features by superimposing convolutional operations in two-dimensional space on different channels,which fails to make good use of temporal information in video.In this paper,an improved model based on convolutional long short-term memory network is proposed.By inserting the temporal codingdecoding phase between the spatial coding phase and the spatial decoding phase,the temporal information is coded and decoded to extract the temporal and spatial motion information better.4.At present,most of the abnormal behavior detection models can only detect the occurrence time of abnormal behavior,but can not be specific to the area where the abnormal behavior occurs.In this paper,a method is proposed to locate the abnormal behavior region from two perspectives of heat map and scatter map,based on the reconstruction error between reconstructed frame and input frame in the model.Through validation on CUHK Avenue dataset,UCSD dataset,Subway dataset and Wuxi Subway dataset,compared with the classical models at home and abroad,the two models proposed in this paper have larger area under Curve(AUC)and smaller equal error rate(EER)on the basis of guaranteeing real-time performance.At the same time,the abnormal area location method can effectively help supervisors to find specific areas where anomalies occur,and improve the practicability of the model.Generally speaking,some useful explorations on the problem of group abnormal behavior detection in subway application scenes are made in this paper,which provided theoretical and methodological support for the practical application.