Review of the Deep Learning Models for Anomaly Detection-Based Video Scrutiny System

Video anomaly detection is a fast-growing computer vision field. Alertness and asset protection explain this. It takes a long time and may yield incorrect results to search surveillance film for illicit activity like a brawl, robbery, or wrong U-turn. High-tech, computer-controlled video surveillance systems are needed for safety. Surveillance cameras regularly capture inexplicable events. This study examines oddities using typical and uncommon video examples. We recommend using the fusion deep learning model to detect abnormalities in weakly tagged training films. Thus, training labels (anomalous or normal) are video-level, not clip-level. This saves people's time from manually marking training video oddities. We use multiple instances learning to automatically develop a deep anomaly ranking model that predicts high anomaly scores for strange video segments. Our method learns a deep anomaly ranking model that predicts high anomaly scores for video segments that deviate from the norm (fusion deep learning model). To help identify anomalies, we add sparsity and temporal smoothness constraints to the ranking loss function during training. We did this to improve data ranking. Their method can be applied in a near-real-time context. Our methodology works for anomaly detection-based video surveillance.