MAAD-GAN: Memory-Augmented Attention-Based Discriminator GAN for Video Anomaly Detection

The detection of anomalies in video data is of great importance in various applications, such as surveillance and industrial monitoring. This paper introduces a novel approach, named MAAD-GAN, for video anomaly detection (VAD) utilizing Generative Adversarial Networks (GANs). The MAAD-GAN framework combines a Wide Residual Network (WRN) in the generator with a memory module to learn the normal patterns present in the training video dataset, enabling the generation of realistic samples. To address the challenge of detecting subtle anomalies and those with motion characteristics, we propose the integration of self-attention in the discriminator model. Our proposed model MAAD-GAN enhances the ability to distinguish between real and generated samples, ensuring that anomalous samples are distorted when reconstructed. Experimental evaluations show the effectiveness of MAAD-GAN as compared to traditional methods on UCSD (University of California, San Diego) Peds2, CUHK Avenue, and ShanghaiTech datasets.