Generalizable Feature Extraction via Transfer Learning in Human Activity Recognition
OPEN ACCESS
Loading...
Author / Producer
Date
2024
Publication Type
Conference Poster
ETH Bibliography
yes
Citations
Altmetric
OPEN ACCESS
Data
Rights / License
Abstract
Efficient human activity recognition (HAR) stands as a critical endeavour across various domains, notably in the medical field where it finds applications in rehabilitation monitoring, disease management, and eldercare. However, the acquisition of datasets in this domain poses significant effort. Current state-of-the-art models often confront limitations stemming from the requirement of extensive data collection.
Recognizing the potential of transfer learning in mitigating these challenges, our study introduces a novel pipeline tailored for transforming kinematic signals associated with human motion into a robust feature representation. Central to this pipeline is a generalized feature extraction neural network capable of discerning salient features from diverse kinematic inputs related to human movement.
Complemented by another neural network responsible for fusion and classification tasks, our research showcases the remarkable performance of this generalized feature extraction pipeline. Across 11 publicly available HAR datasets, our approach outperforms numerous state-of-the-art models. Our comparative analyses against baseline models underscore the superior efficacy of our proposed pipeline.
Our findings highlight the pipeline's exceptional generalization capabilities, even across distinct populations such as those with spinal cord injuries. In essence, our study contributes a significant advancement in HAR, offering a versatile, high-performing solution through the application of transfer learning principles
Permanent link
Publication status
published
External links
Editor
Book title
Journal / series
Volume
Pages / Article No.
Publisher
ETH Zurich, Department of Health Sciences and Technology
Event
Data Science for the Sciences Conference (DS4S 2024)
Edition / version
Methods
Software
Geographic location
Date collected
Date created
Subject
Human activity recognition (HAR); Activities of daily living; Transfer learning; Machine learning; Classification; Neural networks
Organisational unit
03654 - Riener, Robert / Riener, Robert
Notes
This work was partially supported by JST Moonshot R&D, Schweizer Paraplegiker-Stiftung (SPS), and ETH-SPS Digital Transformation in Personalized Health Care for SCI.