Processing math: 0%
Online Transfer Learning With Pseudo Label for Gait Phase Prediction | IEEE Journals & Magazine | IEEE Xplore

Online Transfer Learning With Pseudo Label for Gait Phase Prediction


Abstract:

The gait phase is a fundamental component of lower limb kinematics during normal walking, whose accurate estimation is critical for lower limb rehabilitation, such as con...Show More

Abstract:

The gait phase is a fundamental component of lower limb kinematics during normal walking, whose accurate estimation is critical for lower limb rehabilitation, such as controlling exoskeletons. Most existing deep- and transfer-learning-based methods estimating gait phases generally have an assumption that the target instances are given in advance. However, this assumption may not hold in many practical applications where the target instances not only arrive in an online fashion but also lack ground-truth labels. To address this challenge, we propose a new method called online gait phase predictor (OGPP). Specifically, by designing a new neural network, we first seek to generate a source prototype for source multimodal biomedical signals. The unlabeled target instances are then put into the network to learn a target prototype. Subsequently, a pseudo label generation strategy based on the similarity between the learned source and target prototypes is established to obtain the pseudo target labels. Then we develop a new online transfer learning (OTL) algorithm to train an ensemble classifier by leveraging the source instances, unlabeled target instance, and the pseudo target label. To validate our proposal, an experimental paradigm is designed to collect multimodal biomedical signals, including data from electromyography (EMG), gyroscope, and virtual reality (VR). We further implement OGPP in a real-life scenario to evaluate its real-time performance. Experimental results demonstrate the effectiveness of the proposed method. OGPP achieves a mean accuracy of 79.9% \pm ~3.1 % and time cost of 0.26~\pm ~0.04 s.
Article Sequence Number: 2534415
Date of Publication: 14 October 2024

ISSN Information:

Funding Agency:


References

References is not available for this document.