A Tooled Method for Developing Knowledge-Based Activity Recognizers

Monitoring the daily activities of older adults is a key enabler for aging in place because it reliably indicates whether autonomy is preserved and it prevents unwanted situations (e.g., lack of activity during daytime). To fulfill its promises, activity monitoring requires development methods capable of systematically delivering activity recognizers that are accurate enough to be trusted and accepted by users and their caregivers.This paper presents a systematic approach to developing accurate activity recognizers, based on a tooled method. To achieve accuracy, our strategy is twofold: 1) to encompass the main variations of a target activity by abstracting over descriptions reported by users; 2) to ensure proper customization with respect to user specificities using a dedicated tool. This development method is iterative, allowing to adjust the parameters of an activity recognizer to maximize its accuracy.We validated our approach by applying it to a case study. Specifically, we applied our tooled method to the development of 6 generic activity recognizers, which were then customized with respect to the specificities of 5 older adults, and deployed in their homes during 5 days. Once deployed, the results produced by these activity recognizers were checked daily against activities self-reported by our participants. This experiment shows that 80% of the outputs of our activity detectors were confirmed by the user reports. The accuracy of our approach goes up to 88% when considering the four, more routinized participants.