Video Anomaly Detection Using Convolutional Neural Network

A secure society is necessary for the growth of economy,progress of society and continuous enhancement for national strength.Therefore,intelligent video surveillance technology,which can provide real-time video analysis for anomaly detection,has drawn much attention from researchers in the field of video analysis.Video anomaly detection,as an important branch of intelligent video surveillance technology,focuses on how to extract the relevant features of video for detecting the abnormal event from a given video.Handcrafted features are widely used in traditional video anomaly detection.However,such handcrafted features can not be built without certain prior knowledge,which may vary from scene to scene and can be difficult to define.Different from handcrafted features,features in deep learning can Convolutional Neural Network can be automatically learned relevant during training.Therefore,an algorithm using Convolutional Neural Network for video anomaly detection is proposed in this thesis.Firstly,given the fact that abnormal events are rare in real world and difficult to collect,a video frame prediction based method for video anomaly detection is proposed.A video frame prediction model is an unsupervised algorithm which can extract relevant features for timing signals modeling by learning how to predict normal events.The prediction model is later used for video anomaly detection by computing the prediction error between the predicted frame and the ground truth.An autoencoder structure for feature extraction,timing signals modeling and future frame generation is applied in the model of frame prediction.In this model,3D-CNN and Conv-LSTM are combined to imporve network's capability of modeling timing signals.What'more,fusing different feature maps by concating is also adopted to alleviate information loss brought by feature extraction.Experimental results show that the proposed method does imporve the performance on frame prediction and anomaly detection because it decrease the MSE and increase the PSNR between the predicted result and the ground truth frame.Secondly,a multi-task learning based video prediction network is proposed.Two traditional features and one deep learning feature are adopted to build different auxiliary tasks.One model for auxiliary features,foreground segmentation and optical flow,extracted with traditional methods,MOG2 and Farneback algorithm,is proposed.Another model for optical flow extracted with deep learning method,Flownet2-SD,is also proposed.By predicting features in future freames or by minimizing the feature differences between the predicted future frames and the ground truth frames would help to improve networks capability of predicting future frames and detectiong anomalities.Experimental results show that the auxiliary tasks help to improve performance in two datasets.Experiments carried under the UCSD Ped2 dataset proves that the proposed method achieves higher AUC and lower EER than the state-of-art method,FFPAD.