Lightweight spiking neural networks for human activity recognition on resource-constrained edge wearable devices

Wearable sensor–based human activity recognition is a key area in activity sensing. Spiking neural networks offer a biologically inspired approach for spatiotemporal data processing and are especially suited for time-series prediction. However, existing SNNs mostly use fully connected or 2D convolutional synapses, which cannot fully capture temporal dependencies. Moreover, improving HAR performance remains challenging in terms of computational efficiency and native processing of sparse sensor streams. To address these challenges, we propose SpikSTarS, a novel SNN-based framework for HAR. The framework incorporates a Temporal Response Filter module with temporal convolutions to enhance the spatiotemporal receptive field of synaptic connections, thereby explicitly modeling inter-layer temporal dependencies. In addition, we introduce a STar Aggregate-Redistribute mechanism. Unlike conventional approaches that model channel interactions through distributed structures, the proposed mechanism adopts a centralized aggregation strategy, which improves computational efficiency while reducing sensitivity to the quality of individual sensor channels. Extensive experiments conducted on five public benchmark datasets demonstrate that the proposed SpikSTarS architecture achieves a significant accuracy improvement of 4.37% to 19.39%, while significantly reducing computational complexity. SpikSTarS achieves accuracy competitive with full-precision counterparts on five tested benchmark datasets while substantially cutting computational overhead and shows favourable adaptability across these evaluated data sources.