Semantic Segmentation of Inertial Signals for Human Activity Recognition in Mobile Devices

Sensor-based human activity recognition (HAR) has recently gained significant attention for tasks such as health monitoring, fitness tracking, or elderly care, also thanks to the wide diffusion of wearable devices. In this article, we introduce a solution to perform HAR from signals gathered from inertial measurement units. The proposed approach adapts semantic segmentation techniques, typically used in computer vision, to classify each point of the signal acquired through sensors. The experimental results highlight that the presented methodology can be deployed on mobile hardware platforms, such as smartwatches, while offering significant efficiency in terms of inference time and subsequent energy requirements, as well as achieving good levels of accuracy. In particular, the proposed approach is capable of densely classifying a window of $64 s$ of inertial signals with a Dice score around 0.9, which is only slightly below the F1-score of a state-of-the-art HAR classifier based on non-dense classification (executed on typical $2 s$ long windows), while requiring up to 16 times less power consumption.