Abnormal Behavior Detection In Video Based On Deep Learning

The safety of public places is significant to the construction and development of the country.In the public areas of people's daily life,such as traffic intersections,shopping centers,residential communities,a large number of surveillance cameras are installed for public safety monitoring.Therefore,it has become a highly concerned research direction how to analyze abnormal events of video quickly,real-time,and intelligently.Deep learning has the characteristics of large model capacity and strong adaptability,and has made great achievements in image processing.Based on the knowledge of deep learning,this thesis designs and implements a video abnormal behavior detection network with good performance.First of all,under the premise that normal events are predictable and abnormal events are unpredictable,this thesis proposes the anomaly detection algorithm AP-Crev.AP-Crev predicts possible future results by learning history,and calculates the error between the predicted result and the real result to detect anomalies.In the training phase,AP-Crev uses GAN and unsupervised learning method to predict the future frames of normal events,then obtains the accurate prediction model of normal events.In order to improve the prediction accuracy of future frames,Crev Net is used as the prediction module in AP-Crev.On the one hand,3D-CNN is used in Crev Net to extract more representative spatio-temporal features of historical frames.On the other hand,in Crev Net,the RNN composed of ST-LSTM is used to select the prediction result,which can calculate the dependence of the video frame in the space-temporal dimension.Experiments have proved that AP-Crev improves the accuracy of future frame prediction and also improves the AUC in anomaly detection tasks.Secondly,the data set has the problem of imbalance between normal and abnormal data,which will cause the anomaly detection model to focus on the prediction of the static background when predicting future frames,while ignoring the foreground area where anomalies occur.To solve this,we proposed an anomaly detection network AP-Crev-Att based on the attention mechanism.Specifically,we compute a mask map according to the dataset.After that,we construct an attention map of the dataset during training through the combination of the mask map,background,and optical flow map.Finally,the attention graph is used to give different weights to the background and foreground,which makes the network focus on the prediction of the foreground.The experiment explored the difference of the attention maps obtained by various calculation methods,the influence of attention maps on abnormal scores,and the ability of AP-Crev-Att on anomaly detection tasks.The experimental results validate that the use of the attention mechanism can enhance the anomaly detection performance of the network.In addition,for a specific laboratory scenario,this thesis built a dataset Indoor containing real anomalies,which collects surveillance videos of laboratory seats.Indoor defines six abnormal behaviors which include chatting,falling,throwing objects,turning heads,bowing and lying flat.Indoor collects 10 abnormal videos and 11 normal videos.Through the frame intercepting,screening and marking,the training set and test set are established.The training set contains 7763 normal frames,and the test set contains 5242 abnormal frames.