SSC: Simultaneous Signal Separation and Classification for Concurrent Activities

Wearable sensing systems, such as smartwatch and wristband, can recognize many activities, including sports activities, and hand gestures. Research projects on activity recognition usually assume users perform a single activity. In reality, users can multitask, such as using gestures while running, taking a phone call while tidying up. Existing research on concurrent activity monitoring often requires the user to attach multiple wearables, but attaching multiple devices is often inconvenient and impractical in real-world scenarios. Therefore, monitoring concurrent activities using a single wearable device becomes a challenging task. In this paper, we present a Simultaneous Separation and Classification framework (SSC), which uses a signal separation approach to recognize concurrent activities. The SSC framework includes a Class-Conditioned Signal Tracker module that isolates sensor measurements from different activities in real-time. The isolated signals are used by gesture and gait recognition module to accurately authenticate users in real time while they perform gestures during walking or running, enabling a seamless wearable interface. We tested the proposed solution on a wrist-worn smartwatch. In a study with 10 participants and over 1800 gesture samples, our algorithm achieved 95% accuracy in mobility detection and 93% accuracy in gesture recognition. Furthermore, we could accurately perform user authentication with an accuracy of 91% during walking and 83% while running. Our SSC dataset can be found on Github here.