Big Data-Driven Video Anomaly Detection Using VideoMAE for Visual Analytics in CCTV Surveillance.

Abstract

In the rapidly evolving industry, the need for surveillance is highly needed for the safety of confidential and highly sensitive organizations. With this rise, the demand for efficient security systems has become critical. So, the question is: How to achieve the highest form of security? The obvious answer, which already exists, is a manual monitoring of video footage to check for any suspicious activity. Traditional systems require a lot of manpower as video footage is monitored manually. Imagine a surveillance room full of 50 monitors that display the security footage, and there are just two men observing the footage. How chaotic it is for just two men to manage the whole room and also be vigilant always, to ensure no wrong activity happens. But this isn't very convenient for today's time, it not only overwhelms the observers but also risks critical events being missed due to human limitations. To resolve the issue, we have come up with an intelligent artificial intelligence (AI)-powered system that is capable of spotting anomalous activities in the surveillance footage without any manual effort or stress. An AI-driven system offers seamless, accurate, and proactive monitoring, paving the way for enhanced security in high-stakes environments. We have made use of a deep learning framework for video understanding tasks, which is the Video Masked Autoencoder model, and also a transformer model-Video Vision transformers-for effortless detection of suspicious activity. We have also designed a very simple and attractive web interface where the user can upload the security footage and wait for the model to do its magic. The model detects the anomaly and displays the type of anomaly with the time frame in which the anomaly is occurring. This technique is very effective to find suspicious activities more quickly and efficiently.