CAP-SSM: Context-Aware Prototypes and Selective State-Space Modeling for Efficient Wearable Human Activity Recognition

Human Activity Recognition (HAR) plays a crucial role in Internet of Things (IoT) applications such as healthcare, smart homes, and rehabilitation. Despite remarkable progress, most sensor-based HAR models struggle with subject variability, sensor orientation shifts, and limited computational resources on wearable devices. To overcome these limitations, we propose a novel HAR framework that combines a Convolutional Block Attention Module (CBAM) for adaptive refinement of channel-wise and temporal representations, a light Mamba-based state-space module for efficient modeling of long-range temporal dependencies, and a context-aware prototype classifier. The classifier dynamically adjusts class prototypes according to the input context, yielding adaptive classifications with prototype-based interpretability. Experimental results on five benchmark datasets—UCI-HAR, WISDM, PAMAP2, USC-HAD, and MHEALTH—demonstrate competitive performance comparable to state-of-the-art methods, particularly excelling in efficiency. Our method achieves 99.3% accuracy with 0.48M parameters and approximately 4.6 ms inference latency, demonstrating a superior accuracy–efficiency trade-off compared to recent CNN-, Transformer-, and Mamba-based approaches. We further validate real-time on-device feasibility via Raspberry Pi deployment, confirming practicality on resource-constrained edge platforms. These findings indicate that the proposed model is computationally efficient, robust to inter-subject variability, and well-suited for real-world IoT applications requiring reliable and explainable activity recognition.