Anomaly detection in moving-camera videos with sparse and low-rank matrix decompositions

• Ano de defesa: 2018
• Autor(a) principal: Silva, Eric de Carvalho Jardim
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Federal do Rio de Janeiro; Brasil; Instituto Alberto Luiz Coimbra de Pós-Graduação e Pesquisa de Engenharia; Programa de Pós-Graduação em Engenharia Elétrica; UFRJ
• Palavras-chave em Português: Programação linear; Matrizes esparsas; Detecção de anomalia; Detecção de objetos; Câmeras de vídeo; Segurança; CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA::MEDIDAS ELETRICAS, MAGNETICAS E ELETRONICAS INSTRUMENTACAO::SISTEMAS ELETRONICOS DE MEDIDA E DE CONTROLE
• Link de acesso: http://hdl.handle.net/11422/11668

This work presents two methods based on sparse decompositions that can detect anomalies in video sequences obtained from moving cameras. The first method starts by computing the union of subspaces (UoS) that best represents all the frames from a reference (anomaly-free) video as a low-rank projection plus a sparse residue. Then it performs a low-rank representation of the target (possibly anomalous) video by taking advantage of both the UoS and the sparse residue computed from the reference video. The anomalies are extracted after post-processing this video with these residual data. Such algorithm provides good detection results while at the same time obviating the need for previous video synchronization. However, this technique looses its detection efficiency when target and reference videos presents more severe misalignments. This may happen due to small uncontrolled camera moviment and shaking during the acquisition phase, which is often common in realworld situations. To extend its applicability, a second contribution is proposed in order to cope with these possible pose misalignments. This is done by modeling the target-reference pose discrepancy as geometric transformations acting on the domain of frames of the target video. A complete matrix decomposition algorithm is presented in order to perform a sparse representation of the target video as a sparse combination of the reference video plus a sparse residue, while taking into account the transformation acting on it. Our method is then verified and compared against state-of-the-art techniques using a challenging video dataset, that comprises recordings presenting the described misalignments. Under the evaluation metrics used, the second proposed method exhibits an improvement of at least 16% over the first proposed one, and 22% over the next best rated method.