Transfer learning and data fusion approach to recognize ac8vi8es of daily life

Activity recognition is a core domain within intelligent systems that utilizes the sensing devices available in an environment to identify human activity. Conventional solutions rely on machine-learning approaches and the assumption that the target scenario will Rit the algorithm training conditions, which raises the cost and effort of labelling data, as daily living environments are dynamic, unpredictable, and exposed to new activities. Hence, we take advantage of the ubiquitous presence of personal gadgets such as smart-watches combined with data fusion approaches to dynamically transfer learned knowledge across devices in a natural environment while performing daily living activities. In this paper, we focus on recognizing walking as an activity, which might enable carers or medical practitioners to monitor the risk of falling or suffering from a chronic disease whose progression is linked to a reduction in movement and mobility. Preliminary results show a 2% increase in activity recognition accuracy on the wearable approach, and a 10% improvement in accuracy when combining features from both wearable and environmental domains.