Pelvic IMU-Based Gait Phase Classification Using Kernel-Embedded ELM: A Preliminary Validation Framework for Future Hemiplegic Gait Applications

초록

This study proposes a geometry-aware gait phase classification framework designed for real-time application using a single inertial measurement unit (IMU) placed at the center of the pelvis. Our approach employs a relevant classification scheme consisting of three gait combinations: (1) left swing with right stance, (2) right swing with left stance, and (3) double stance. To extract meaningful temporal dynamics from a single IMU, raw signals are encoded into 6 × 6 covariance matrices, which are then embedded into a nonlinear feature space using a Gaussian kernel. This embedding captures underlying gait structure while respecting the geometry of the SPD manifold. A lightweight Extreme Learning Machine (ELM) classifier processes the resulting kernel vectors for efficient phase recognition. The proposed system was validated on data from eight healthy participants using four different geometry-aware distance metrics. Among them, the Log-Euclidean distance yielded the highest average classification accuracy of 89.18 % pm 1.51 % across five trials. These results demonstrate that even with a single pelvic IMU, our method can reliably distinguish bilateral gait phase dynamics, highlighting its potential for future application in hemiplegic gait assessment and wearable assistive robotics. © 2025 ICROS.

키워드