Surveillance Video Anomaly Detection With Multi-Branch Gan

Abstract

Identifying irregularities in surveillance videos is essential for maintaining public safety in various settings. Conventional methods, which often depend on human oversight, can be inefficient and susceptible to errors. Nevertheless, with the advancement of deep learning technologies, the task of detecting anomalies in real-time can now be automated. Generative Adversarial Networks (GANs) have demonstrated their effectiveness in video analysis, enabling the automated identification of anomalies. This project introduces an innovative approach that employs a Multi-Branch GAN (M-GAN) model specifically crafted for the detection of abnormal events in surveillance footage. The M-GAN utilizes a two-phase approach: it initially learns the patterns of typical activities and subsequently identifies any deviations as potential anomalies. A key advantage of this model is its ability to function without requiring labelled anomaly data, which increases its flexibility across different environments and conditions. Experimental results demonstrate that the M-GAN outperforms conventional GAN-based methods, achieving superior precision and recall rates in detecting abnormal occurrences. This outstanding performance positions M-GAN as an excellent solution for instantaneousabnormal event detection in surveillance systems, enhancing safety and security while reducing dependence on manual monitoring.