BROWSE
Detail
Conference Paper
IMU2IMG: IMU in the Language of Vision Foundation Models
Cited 0 time in 
Share 
Share
- Authors
- Sun Kyung Lee, Se Won Oh, Gyuwon Jung, Hyuntae Jeong, Seungeun Chung, Jeong Mook Lim, Kyoung Ju Noh
- Issue Date
- 2025-10
- Citation
- International Conference on Pervasive and Ubiquitous Computing (UbiComp) 2025, pp.1-6
- Language
- English
- Type
- Conference Paper
- Abstract
- Motivated by the growing interest in foundation models across domains such as vision and language, this study aims to explore their potential in the field of human locomotion recognition using wearable sensor data. In response to Task 1 of Sussex-Huawei Locomotion-Transportation (SHL) recognition challenge 2025, which focuses on the foundation model based user-independent activity recognition with inertial signals, our team 'HELP' proposes IMU2IMG, an algorithmic pipeline that leverages a vision foundation model for prediction. Specifically, our method transforms multimodal inertial data into RGB images, enabling explicit alignment between sensor representations and visual tokens. Through extensive experiments, we compare IMU2IMG with conventional machine learning, deep learning, and recent foundation model-based baselines. Our results demonstrate superior performance and offer insights on how sensor-to-vision mappings can support robust and interpretable human activity recognition.
- KSP Keywords
- Human activity recognition, Human locomotion, Inertial Data, RGB image, Wearable sensor data, deep learning(DL), machine Learning, model-based, superior performance