Efficient Sensing and Classification for Extended Battery Life

The proliferation of wearable technologies has facilitated continuous health monitoring in daily life. However, the individualized nature of human movement and the need for low-power consumption present challenges for these systems. Traditionally, machine learning methods used in these wearables have fixed properties, such as sensor sampling rate, which overlook the necessity for personalized computational algorithms. Addressing this, we propose a resource-efficient, real-time human activity recognition framework, transforming the multi-class classification problem into a hierarchical model based on the Metabolic Equivalent of Task (MET), creating a personalized structure for each individual. Our new configurable classification paradigm is designed to be energy- and memory-efficient, considering the limited resources of wearable devices. The results suggest our proposed system accurately detects activities in different personalized scenarios between 94.5% and 96.9% of the time using limited memory while reducing power consumption by up to 17.2% compared to conventional methods.