DG-GAN: A Deep Neural Network for Real-World Anomaly Detection in Surveillance Videos

Identification of anomalies in real-time surveillance videos is crucial for ensuring the security and safety of various environments, such as public spaces, critical infrastructure, and commercial facilities. Automated anomaly detection plays an important role in surveillance systems as it reduces the need for human involvement and the associated costs. Autoencoders in recent years have proven to be effective as anomaly detectors learn only from normal data. Furthermore, GANs have been used to provide additional training instances to classifiers, increasing their accuracy and robustness. While GAN has achieved remarkable performance in generating realistic images, there exist major issues in training GAN such as mode collapse, non-convergence, and instability. These challenges often arise due to improper design of network architecture, selection of optimization algorithms, and use of objective functions. To address these challenges, we present a novel approach called the Dual Generator-based Generative Adversarial Network (DG-GAN). This network comprises two distinct components: a temporal generator and an image generator. The former accepts a single input in the form of a latent variable and generates a series of latent variables while the latter processes this sequence of latent variables, transforming them into a complete video sequence. Our proposed framework was evaluated using three data sets: UCSD Ped2, ShanghaiTech Campus, and CUHK Avenue. The experimental results demonstrate that our framework is superior to existing learning-based approaches in terms of AUC, achieving scores of 98.08%, 88.43%, and 92.2% respectively.