Spatiotemporal Graph Neural Network for Lower-Limb Activity Recognition Using IMUs

Abstract

Wearable inertial measurement units (IMUs) provide a practical and privacy-preserving sensing modality for lower-limb activity recognition in assistive robotics, rehabilitation monitoring, and movement analysis. However, many IMUbased human activity recognition methods model multichannel sensor streams mainly as flat temporal sequences, limiting their ability to exploit the structured dependencies among bilateral lower-limb segments. In this work, we investigate a graph-based activity recognition framework that reorganizes wearable IMU windows into a fixed six-node lower-limb motion-chain graph and uses a spatiotemporal graph convolutional network (ST-GCN) to model anatomical spatial dependencies. A parallel cumulative gated recurrent unit (cGRU) captures temporal evolution, and self-attention fuses temporal and graph-based features for recognition, with short-horizon prediction used only as an auxiliary regularizer. Under leave-one-subject-out crossvalidation (LOOCV), the proposed method achieved 96.74% accuracy and 96.28% macro F1 on HuGaDB-3, and 87.68% accuracy and $82.57 \%$ macro F1 on a self-collected lower-limb IMU dataset (SC-LL). These results, obtained on two small subject cohorts, suggest that anatomically structured graph modeling can improve wearable IMU-based activity recognition, particularly when the sensor configuration is well aligned with the assumed lower-limb graph topology.