Deep learning-based anomaly detection in video surveillance

In this day and age of smart cities, the use of video monitoring has assumed a position of critical significance. Large numbers of surveillance cameras have been installed in public and private locations for the purpose of monitoring the properties of infrastructure and ensuring the safety of the general public. This research provides a multi-modal CNN-BiLSTM autoencoder framework for detecting anomalous events in important surveillance environments such as bank ATMs. The approach is built on semi-supervised deep learning and uses multi-modal data. In addition, because there was no publicly accessible dataset for ATM surveillance, we created a one-of-a-kind RGB+D dataset specifically for this purpose. This was done because there was no dataset for ATM surveillance in the public domain. The proposed methodology is validated by testing it on the RGB+D dataset that was collected as well as two other real-world benchmark video anomaly datasets: Avenue and UCFCrime2Local.