The Research And Implement Of The Crowd Abnormal Event Detection Methods Based On The Video Surveillance

The timely discovery of abnormal events in public places will help relevant departments to respond and rescue in a timely manner,reducing casualties and property losses.With the popularization and development of video surveillance,real-time anomaly detection and early warning through intelligent video analysis have become an important means.Under the influence of complex scene,noise and other factors,the existing video-based anomaly detection methods are prone to omission and false alarm.Therefore,improving the adaptability of video-based anomaly detection methods in complex scenarios has become an important research direction.This dissertation mainly studies and implements the detection of crowd abnormal events based on video,proposes two methods suitable for the detection of global and local abnormal events,and designs and implements the crowd abnormal events detection software based on video analysis,which can detect the crowd abnormal events in the video surveillance.The specific work is as follows:In terms of detecting local abnormal events,an improved deep learning model of convolutional autoencoder is proposed.This model is an unsupervised learning model,which adds a LSTM network on the basis of traditional convolutional autoencoder.The model can extract the appearance features of normal crowd movement by performing convolution.Through the LSTM network,the temporal correlation of appearance features is captured,and the deep spatio-temporal feature extraction of normal crowd movement is realized.The reconstruction error of the model is used as the basis of anomaly judgment to construct the anomaly scoring function.Through simulation experiments on Avenue dataset,UCSD ped1 and ped2 datasets,AUC values detected in real time reached 84.20%,84.65% and 89.32%,and EER values were 22.87%,22.80% and 20.74%,respectively.Based on the performance of the three datasets,this method has strong adaptability in different scenarios.In terms of global abnormal event detection,a method based on foreground extraction and optical flow features was proposed.ViBe foreground object extraction algorithm is used to calculate the proportion of foreground objects and measure the size of the crowd.Optical flow features are obtained by HS algorithm,and the intensity of crowd movement is analyzed.An anomaly scoring function about the proportion of foreground objects and optical flow velocity is constructed,and the threshold is set for abnormality determination.Based on the performance of the dataset,this method has strong adaptability in different scenarios.In terms of the design and implementation of the software for abnormal event detection based on video,Eric 6.1.7 is used as the development platform,and PyQt 5 is adopted to design the interface.By analyzing the relevant functional requirements,the software includes three modules: ‘user login',‘local abnormal event detection' and ‘global abnormal event detection'.It has the functions of restricting user permissions,crowd feature extraction and anomaly event detection.