Robust approaches for anomaly detection applied to video surveillance

Detalhes bibliográficos
Ano de defesa: 2018
Autor(a) principal: Rensso Victor Hugo Mora Colque
Orientador(a): Não Informado pela instituição
Banca de defesa: Não Informado pela instituição
Tipo de documento: Tese
Tipo de acesso: Acesso aberto
Idioma: eng
Instituição de defesa: Universidade Federal de Minas Gerais
Brasil
ICX - DEPARTAMENTO DE CIÊNCIA DA COMPUTAÇÃO
Programa de Pós-Graduação em Ciência da Computação
UFMG
Programa de Pós-Graduação: Não Informado pela instituição
Departamento: Não Informado pela instituição
País: Não Informado pela instituição
Palavras-chave em Português:
Link de acesso: http://hdl.handle.net/1843/32271
Resumo: Modeling human behavior and activity patterns for detection of anomalous events has attracted significant research interest in recent years, particularly among the video surveillance community. An anomalous event might be characterized by the deviation from the normal or usual, but not necessarily in an undesirable manner. One of the main challenges of detecting such events is the difficulty to create models due to their unpredictability and their dependency on the context of the scene. Anomalous events detection or anomaly recognition for surveillance videos is a very hard problem. Since anomalous events depend on the characteristic or the context of a specific scene. Although many contexts could be similar, the events that can be considered anomalous are also infinity, i.e., cannot be learned beforehand. In this dissertation, we propose three approaches to detect anomalous patterns in surveillance video sequences. In the first approach, we present an approach based on a handcrafted feature descriptor that employs general concepts, such as orientation, velocity, and entropy to build a descriptor for spatiotemporal regions. With this histogram, we can compare them and detect anomalies in video sequences. The main advantage of this approach is its simplicity and promising results that will be show in the experimental results, where our descriptors had well performance in famous dataset as UCSD and Subway, reaching comparative results with the estate of the art, specially in UCSD peds2 view. This results show that this model fits well in scenes with crowds. In the second proposal, we develop an approach based on human-object interactions. This approach explores the scene context to determine normal patterns and finally detect whether some video segment contains a possible anomalous event. To validate this approach we proposed a novel dataset which contains anomalies based on the human object interactions, the results are promising, however, this approach must be extended to be robust to more situations and environments. In the third approach, we propose a novel method based on semantic information of people movement. While, most studies focus in information extracted from spatiotemporal regions, our approach detects anomalies based on human trajectory. The results show that our model is suitable to detect anomalies in environments where trajectory of the people could be extracted. The main difference among the proposed approaches is the source to describe the events in the scene. The first method intends to represent the scene from spatiotemporal regions, the second uses the human-object interactions and the third uses the people trajectory. Each approach is oriented to certain anomaly types, having advantages and disadvantages according to the inherit limitation of the source and to the subjective of normal and anomaly event definition in a determinate context.