Learning Human Behaviour Patterns by Trajectory and Activity Recognition

• Autor(a) principal: Fernandes, Letícia Maria Sousa
• Data de Publicação: 2019
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Texto Completo: http://hdl.handle.net/10362/87075
• topic: Human Behaviour; Pattern Recognition; Anomaly Detection; Ambient Assisted Living; Probability Density Function; Clustering; Domínio/Área Científica::Engenharia e Tecnologia::Outras Engenharias e Tecnologias

The world’s population is ageing, increasing the awareness of neurological and behavioural impairments that may arise from the human ageing. These impairments can be manifested by cognitive conditions or mobility reduction. These conditions are difficult to be detected on time, relying only on the periodic medical appointments. Therefore, there is a lack of routine screening which demands the development of solutions to better assist and monitor human behaviour. The available technologies to monitor human behaviour are limited to indoors and require the installation of sensors around the user’s homes presenting high maintenance and installation costs. With the widespread use of smartphones, it is possible to take advantage of their sensing information to better assist the elderly population. This study investigates the question of what we can learn about human pattern behaviour from this rich and pervasive mobile sensing data. A deployment of a data collection over a period of 6 months was designed to measure three different human routines through human trajectory analysis and activity recognition comprising indoor and outdoor environment. A framework for modelling human behaviour was developed using human motion features, extracted in an unsupervised and supervised manner. The unsupervised feature extraction is able to measure mobility properties such as step length estimation, user points of interest or even locomotion activities inferred from an user-independent trained classifier. The supervised feature extraction was design to be user-dependent as each user may have specific behaviours that are common to his/her routine. The human patterns were modelled through probability density functions and clustering approaches. Using the human learned patterns, inferences about the current human behaviour were continuously quantified by an anomaly detection algorithm, where distance measurements were used to detect significant changes in behaviour. Experimental results demonstrate the effectiveness of the proposed framework that revealed an increase potential to learn behaviour patterns and detect anomalies.