User-trained activity recognition using smartphones and weak supervision

With many people now carrying some form of smart device, the feasibility of capturing movement data has never been so practical. Current methods of activity recognition use complex sensor arrangements or potentially biased questionnaires. This paper presents a method of activity recognition which is trained by the user on their own device, reducing the requirement for laboratory-based data capture experiments. Requests will be made to the user to provide labelling information at fixed intervals and a classifier is trained using these user-captured labels. These requests will come in the form of notifications on their smart device. Many activity recognition systems are limited to capturing the activities they are pre-trained with. However, allowing the user to provide their own sparse labels provides an opportunity to capture a larger range of activities. The novel contributions of this paper are in the combination of experience sampling with multiple instance learning and a clustering method to provide a simpler method of data capture for activity recognition.