Low-Cost Indoor Localisation Based on Inertial Sensors, Wi-Fi and Sound

• Autor(a) principal: Teófilo, Ana Filipa Frazão de Almeida
• Data de Publicação: 2021
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10362/130418
• topic: Ambient Assisted Living; Indoor location; Machine Learning; Deep Learning; Data Fusion; Domínio/Área Científica::Engenharia e Tecnologia::Outras Engenharias e Tecnologias

The average life expectancy has been increasing in the last decades, creating the need for new technologies to improve the quality of life of the elderly. In the Ambient Assisted Living scope, indoor location systems emerged as a promising technology capable of sup porting the elderly, providing them a safer environment to live in, and promoting their autonomy. Current indoor location technologies are divided into two categories, depend ing on their need for additional infrastructure. Infrastructure-based solutions require expensive deployment and maintenance. On the other hand, most infrastructure-free systems rely on a single source of information, being highly dependent on its availability. Such systems will hardly be deployed in real-life scenarios, as they cannot handle the absence of their source of information. An efficient solution must, thus, guarantee the continuous indoor positioning of the elderly. This work proposes a new room-level low-cost indoor location algorithm. It relies on three information sources: inertial sensors, to reconstruct users’ trajectories; environ mental sound, to exploit the unique characteristics of each home division; and Wi-Fi, to estimate the distance to the Access Point in the neighbourhood. Two data collection protocols were designed to resemble a real living scenario, and a data processing stage was applied to the collected data. Then, each source was used to train individual Ma chine Learning (including Deep Learning) algorithms to identify room-level positions. As each source provides different information to the classification, the data were merged to produce a more robust localization. Three data fusion approaches (input-level, early, and late fusion) were implemented for this goal, providing a final output containing complementary contributions from all data sources. Experimental results show that the performance improved when more than one source was used, attaining a weighted F1-score of 81.8% in the localization between seven home divisions. In conclusion, the evaluation of the developed algorithm shows that it can achieve accurate room-level indoor localization, being, thus, suitable to be applied in Ambient Assisted Living scenarios.