Deep Learning-Based Smart Surveillance System

Abstract

The role of CCTV cameras has been overgrown in this generation. CCTV cameras are installed all over the places for surveillance and security. Many surveillance systems still require human supervision. Recent advances in computer vision are, thus, seen as an important trend in video surveillance that could lead to dramatic efficiency gains. Various public places like shopping malls, supermarkets, ATMs, banks, and other places, where CCTV cameras are available, are the places we should concentrate on. Security can be characterized in various terms in various settings like robbery distinguishing proof, brutality recognition, odds of a blast, and so on. In jam-packed public places, the term security covers practically a wide range of strange occasions. So, it is important and challenging to build a model which detects these abnormal activities and generates some kind of alert. We used a combination of convolutional neural networks (CNNs) and long short-term memory (LSTM) which involve the concept of deep neural networks. It extracts the spatial–temporal features of the images and calculates the Euclidean distance between the original and reconstructed batch of images. We converted the training videos into images to train the model and calculated the loss between the images to identify the abnormality. To validate the proposed algorithm, 4 datasets as HOLLYWOOD, UCF101, HMDB51, and WEIZMANN are used for action recognition. The proposed technique performs better than the existing one. We made use of Jupiter notebook and Python frameworks.