Video Anomaly Detection using Deep Autoencoder and Generative Adversarial Network

Abstract

Abnormal event detection in videos refers to identifying events that deviate from anticipated newlinepatterns. The high dimensionality of video data poses substantial complexity and newlinechallenges in detecting abnormal patterns within large-scale datasets. Consequently, the newlinesignificance of unsupervised anomaly detection methods has emerged prominently in video newlinesurveillance systems. The unsupervised techniques are becoming more significant because newlineannotating extensive data is time-consuming and burdensome. The video anomaly detection newlinetechniques are broadly classified into frame reconstruction and future frame prediction-based newlinemethods. In reconstruction-based approaches, the model learns to recreate the input data with newlineminimal error. Moreover, during testing, the higher reconstruction error of the pixel signifies newlinean abnormal event. In contrast, the prediction-based approach learns to predict the future newlineframe of input video data based on past sequences, and the higher prediction error indicates newlinethe anomalies. Due to the high generalization capability of the deep autoencoders, sometimes, newlinethe reconstruction-based approaches recreate the abnormal patterns along with the newlinenormal ones, leading to false detection. Therefore the frame prediction-based methods are newlinemore famous. In this thesis, we have designed both the reconstruction and prediction-based newlinedeep neural network to learn the typical patterns of the video data to detect anomalies. newlineFirstly, we have developed a CNN-LSTM-based reconstructive autoencoder network, newlinenamed inter-fused autoencoder (IFA), to model the regular pattern of video data. During newlinetesting, the higher reconstruction error signifies abnormal events in a video. For a seq2seq newlinemodel, the decoder output mainly depends on the context of the last hidden state of the newlineencoder; therefore, there is a high probability of losing the initial data for a long sequence. newlineTo address this, we have employed a multiplicative attention block on the decoder side of the newlinereconstructive autoencoder network, named deep multiplicative atten