CASNN: Continuous Adaptive SNN for Human Activity Recognition

Human activity recognition (HAR) is an application of great importance in the Internet of Things (IoT). Inertial measurement units (IMU) on wearable devices provide the primary source of time-series data for HAR. This paper focuses on addressing the energy consumption and performance issues in real-world scenarios for continuous HAR using time-series signals. We present a continuous adaptive spiking neural network (CASNN) suitable for low-power wearable devices. CASNN is implemented by introducing early exit into SNN, and the early exit branches do not require additional classifiers. The results show that CASNN can reduce over 56% FLOPs and improve accuracy by 4% on average across two datasets, through additional cross-person generalization capabilities on the continuous HAR dataset.