ML-Driven User Activity-Based GNSS Activation for Power Optimization in Resource-Constrained Environments

Abstract

The aging population represents an increasing burden on healthcare systems, which is shifting policies from institutionalization to aging in the community. Remote monitoring offers efficient solutions that bridge the gaps between healthcare and where elderly people really want to live every day. However, the adoption of such systems remains low, especially in resource-constrained environments like underdeveloped regions and rural areas, due to the lack of resources often taken for granted in system design. Location is one of the main types of information to monitor, as it provides information about behavior and physical activity. Global Navigation Satellite System (GNSS) is the de facto technology, and although its high-power consumption aligns poorly with battery-powered devices, it is still the best choice for accurate and reliable remote localization of pedestrians. Deciding when to turn on/off the GNSS receiver based on context is a key strategy for power optimization, the two main types of contexts being the user’s position and activity. However, existing methods in the literature are not suitable for resource-constrained environments because they require the installation of beacons, which entail additional cost and power consumption, or assume the availability of external signals that are not met in such environments, or are based on simple user activity detection. This work proposes a new GNSS activation method based on detecting the specific walking activity for changing locations. In resource-constrained rural environments, people typically spend most of their time outdoors near their houses, where it is not necessary to activate the GNSS so frequently to monitor them. Restricting the GNSS activation to the moments in which they are moving to a different location could be enough and would reduce the power consumption. Four machine learning (ML) classification models [long short-term memory (LSTM), extreme gradient boosting (XGBoost), support vector machine (SVM), and random forest (RF)] have been implemented and evaluated using a smartwatch’s inertial sensor data. The best model, XGBoost, was exported to a custom-designed embedded system and evaluated in real-world tests. It demonstrated over 40% power savings compared to conventional motion-based methods.