A knowledge-light approach to personalised and open-ended human activity recognition

Abstract Human Activity Recognition (HAR) is a core component of clinical decision support systems that rely on activity monitoring for self-management of chronic conditions such as Musculoskeletal Disorders. Deployment success of such applications in part depend on their ability to adapt to individual variations in human movement and to facilitate a range of human activity classes. Research in personalised HAR aims to learn models that are sensitive to the subtle nuances in human movement whilst Open-ended HAR learns models that can recognise activity classes out of the pre-defined set available at training. Current approaches to personalised HAR impose a data collection burden on the end user; whilst Open-ended HAR algorithms are heavily reliant on intermediary-level class descriptions. Instead of these “knowledge-intensive” HAR algorithms; in this article, we propose a “knowledge-light” method. Specifically, we show how by using a few seconds of raw sensor data, obtained through micro-interactions with the end-user, we can effectively personalise HAR models and transfer recognition functionality to new activities with zero re-training of the model after deployment. We introduce a Personalised Open-ended HAR algorithm, MN Z , a user context aware Matching Network architecture and evaluate on 3 HAR data sources. Performance results show up to 48.9% improvement with personalisation and up to 18.3% improvement compared to the most common “knowledge-intensive” Open-ended HAR algorithms.