Intelligent Eye: Anomaly Detection in surveillance system

Abstract

In this work, we propose the Intelligent Eye—an advanced surveillance video stream anomaly detection system using modern, state-of-the-art deep learning architectures, including EfficientNet and Vision Transformer (ViT) for this task. With increased demand for effective surveillance, the identification of unusual events becomes imminent—automatically detecting instances of violence, unauthorized entry, or safety threats. The truth is that the traditional ways of anomaly detection have, for most parts, failed to live up to the expectation on complexity and volume of today's surveillance data. We have also used Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM) networks, for processing thermal imagery in cases of detected events in outdoor environments, due to the challenges of data in thermal video that usually are not visible in standard surveillance video. We assembled a new dataset by combining standard and thermal imagery to train and validate our models, hence enabling robust performance for diverse scenarios which includes violence, unauthorized wall crossing, human fall and, fire and smoke. The results obtained from our system showed an accurate anomaly detection over visual and thermal images, which indicates our integrated deep learning approach works effectively. This work contributed to the progress of the theory in anomaly detection and, at the same time, for practical solutions that enhance security and safety measures executed by surveillance systems in their applications.