3D-CLSTM-GAN: Three-Dimensional Conventional Long Short-Term Memory Generative Adversarial Network for Anomaly Detection in Video Surveillance

Abstract

Anomaly detection in videos is a challenging task due to unbounded nature of anomalies in different contexts. Supervised learning is not ideal for such computer vision applications due uncertainty and abnormality definition. Generative Adversarial Network (GAN) emerged as a class of neural networks with unsupervised learning. In this paper, we proposed a GAN architecture named 3D-CLSTM-GAN for real time detection of anomalies in surveillance videos. We used an 3D convolutions and Conventional Long Short Term Memory (ConvLSTM) in the generator for efficient learning of spatio-temporal features. We used an encoder based discriminator which discriminates generated video frames. We used a loss function that promotes robustness in anomaly detection due to its ability for quick convergence of the underlying model. After learning from normal data distribution, the proposed architecture detects anomalies based reconstruction error which reflects deviation of frames representation. The proposed 3D-CLSTM-GAN is lightweight due to the 3D ConLSTM which exploits 3D convolutions efficiently. The 3D-CLSTM-GAN architecture is evaluated with empirical study using three benchmark datasets such as ShanghaiTech, CHUK Avenue and UCSD. Experimental results revealed that the proposed 3D-CLSTM-GAN outperforms the state of the art models such as PPC+SFA, Conv-AE, ConvLSTM-AE and TSC.