Unusual Event Detection in Surveillance Videos

• Autor(a) principal: Nazaré, Tiago Santana de
• Data de Publicação: 2020
• Tipo de documento: Tese
• Idioma: eng
• Título da fonte: Biblioteca Digital de Teses e Dissertações da USP
• Texto Completo: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-16022021-150800/
• topic: Anomaly detection; Computer vision; Detecção de anomalias; Fluxo ótico; Optical-flow; Surveillance videos; Vídeos de segurança; Visão computacional

Presently, surveillance cameras have been massively employed to monitor public spaces such as malls, train stations and airports. The video feed generated by several of these security cameras is monitored, in real-time, by a small group of people in a control room to detect anomalous behaviors. Nonetheless, human monitoring is extremely ineffective when it comes to detecting anomalies in surveillance footage, mainly because such task is both tedious (most of the time nothing interesting/abnormal happens) and challenging (a single person is in responsible for keeping track of multiple cameras at the same time). The aforementioned problems motivated the machine vision community to develop automated methods for detecting unusual behaviors in security videos. Despite recent advances in this area, we have noticed that current video anomaly detection methods have some gaps regarding: i) the lack of noise removal/management techniques when modeling motion patterns using optical-flow estimates; and ii) the need for a more adaptive approach to tackle changes in viewing distances. Motivated by those issues, we proposed some methods/studies aiming at improving anomaly detection in surveillance videos, while maintaining (or reducing) computational cost. Our experiments show that employing lightweight filtering to optical-flow estimates and anomaly scores can significantly improve anomaly detection performance in surveillance scenarios, without increasing computational complexity. Furthermore, we presented an automatic method to estimate changes in object size caused by variations in viewing changes, which is capable of improving anomaly detection performance and reducing setup time. Based on those findings, we designed an anomaly detection method that is capable to achieve state-of-the-art anomaly detection performance in challenging surveillance scenarios employing only optical-flow information. Additionally, we showed that a domain-specific auto-encoder is capable of achieving comparable anomaly detection results to the ones of features from pre-trained CNN models, while having a significant lower computational complexity (smaller number of network parameters).