Anomaly Detection in Surveillance Videos Using Hybrid Deep Learning Model DBNSSGAN

Abstract

Urban planners and academics are influenced by the contemporary notion of smart cities to create modern, secure, and sustainable infrastructure that offers a respectable standard of living to its inhabitants. In order to improve citizen safety and well-being, video surveillance cameras have been installed to meet this demand. Even with today's scientific advancements in technology, abnormal event detection in CCTV footage and surveillance video remains difficult and time-consuming for humans to complete. Surveillance videos that contain anomalous events are automatically identified by video anomaly detection. The ability to determine whether a video contains anomalous events has improved in previous efforts. Since the development of deep learning methods, researchers have become interested in automatic video surveillance. The task of video anomaly detection, can be approached as a semi-supervised learning problem because to the strong bias in the datasets towards normal samples. The widely used reconstruction techniques solely use regular images to train the network. Assuming that the network cannot precisely recreate anomalous regions, these approaches identify anomalous occurrences by comparing the input with the reconstructed image. These approaches, however, have a significant drawback in that the anomaly zones are not sufficiently generic. This issue narrows the difference between the reconstructed and anomalous input images, which decreases the capacity to detect anomalies. In this paper the semi supervised Generative Adversarial Networks (SSGAN) is combined with Deep belief network (DBN) in detecting the abnormal events in surveillance video which greatly improves the quality of reconstruction and classifies the anomaly effectively. The outcomes are compared with the most advanced deep learning methods using two well-known surveillance data sets.