Exploring radar sensing for gesture recognition

• Autor(a) principal: Santana, Luís Fernando do Vale
• 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/10773/33726
• topic: Radar; Aphasia; Gesture; Communication; Gesture recognition; Smart home; Transfer learning; Non-intrusive sensors

Communication disorders have a notable negative impact on people’s lives, leading to isolation, depression and loss of independence. Over the years, many different approaches to attenuate these problems were proposed, although most come with noticeable drawbacks. Lack of versatility, intrusive solutions or the need to carry a device around are some of the problems that these solutions encounter. Radars have seen an increase in use over the past few years and even spreading to different areas such as the automotive and health sectors. This technology is non-intrusive, not sensitive to changes in environmental conditions such as lighting, and does not intrude on the user’s privacy unlike cameras. In this dissertation and in the scope of the APH-ALARM project, the author tests the radar in a gesture recognition context to support communication in the bedroom scenario. In this scenario, the user is someone with communication problems, lying in their bed trying to communicate with a family member inside or outside the house. The use of gestures allows the user to have assistance communicating and helps express their wants or needs. To recognize the gestures executed by the user, it is necessary to capture the movement. To demonstrate the capabilities of the technology, a proof of concept system was implemented, which captures the data, filters and transforms it into images used as input for a gesture classification model. To evaluate the solution, we recorded ten repetitions of five arm gestures executed by four people. A subject independent solution proved to be more challenging when compared to a subject dependent solution, where all datasets but one achieved a median accuracy above 70% with most going over 90%.