Real time Video Anomaly Detection using Deep Belief Network with Semi Supervised GAN

Abstract

In recent years, real-time video anomaly detection has emerged as a crucial task in various applications, including surveillance, security, and smart cities. Big data is created by video streams that are recorded as the number of surveillance cameras increases. It has become imperative to analyze the video streams gathered from those traffic surveillance cameras in order to identify unusual, suspicious events and various harmful activities, as it is not feasible to view, evaluate, and understand the contents of these films with human labor.The article introduced the Deep Belief Network (DBN) in conjunction with a Semi-Supervised Generative Adversarial Network (GAN) to effectively detect anomalies in video streams. Our method utilizes the unsupervised learning capabilities of DBNs to capture complex patterns from normal video frames while the semi-supervised GAN aids in generating realistic examples of both normal and anomalous behaviors. By integrating these two frameworks, we achieve improved feature extraction and robust anomaly detection. Our approach is evaluated on real time datasets collected from the CCTV footages from news and demonstrate significant improvements in detection accuracy and compared to existing methods. Our findings suggest that this hybrid model is a good alternative for dynamic environments where anomalies may regularly occur because it not only improves performance but also adapts well to real-time applications.