Evaluation of spatial-temporal anomalies in the analysis of human movement

• Autor(a) principal: Varandas, Rui Pedro Sousa
• Data de Publicação: 2018
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10362/58215
• topic: Time Series; Anomaly Detection; Human Motion; Unsupervised Learning; Manufacture; Domínio/Área Científica::Ciências Médicas::Outras Ciências Médicas

The dissemination of Internet of Things solutions, such as smartphones, lead to the appearance of devices that allow to monitor the activities of their users. In manufacture, the performed tasks consist on sets of predetermined movements that are exhaustively repeated, forming a repetitive behaviour. Additionally, there are planned and unplanned events on manufacturing production lines which cause the repetitive behaviour to stop. The execution of improper movements and the existence of events that might prejudice the productive system are regarded as anomalies. In this work, it was investigated the feasibility of the evaluation of spatial-temporal anomaly detection in the analysis of human movement. It is proposed a framework capable of detecting anomalies in generic repetitive time series, thus being adequate to handle Human motion from industrial scenarios. The proposed framework consists of (1) a new unsupervised segmentation algorithm; (2) feature extraction, selection and dimensionality reduction; (3) unsupervised classification based on DBSCAN used to distinguish normal and anomalous instances. The proposed solution was applied in four different datasets. Two of those datasets were synthetic and two were composed of real-world data, namely, electrocardiography data and human movement in manufacture. The yielded results demonstrated not only that anomaly detection in human motion is possible, but that the developed framework is generic and, with examples, it was shown that it may be applied in general repetitive time series with little adaptation effort for different domains. The results showed that the proposed framework has the potential to be applied in manufacturing production lines to monitor the employees movements, acting as a tool to detect both planned and unplanned events, and ultimately reduce the risk of appearance of musculoskeletal disorders in industrial settings in long-term.