Text2IMU: Advancing Human Activity Recognition by Text-Driven IMU Data Synthesis

Lars Ole Haeusler, Lena Uhlenberg, Oliver Amft

Confirmation: I have read and agree with the IEEE BSN 2025 conference submission's policy on behalf of myself and my co-authors.

Keywords: Synthetic Data, Human Activity Recognition, IMU sensor synthesis, Generative Motion Synthesis

TL;DR: We introduce Text2IMU: A framework to generate realistic synthetic Inertial Measurement Unit data from textual descriptions that can be used to train Human Activity Recognition models solely on synthetic data.

Abstract: Inspired by the progress of motion synthesis models, we leverage cross-modality transfer to generate realistic synthetic Inertial Measurement Unit (IMU) data from textual descriptions, hence Text2IMU. We use an established motion synthesis model and textual descriptions to generate sequences of 3D human activities. To obtain realistic and diverse sensor readings, we created multiple body surface models with different body morphologies. With the text prompts, we let the surface models perform activities and synthesise acceleration and gyroscope data for multiple virtual IMU positions. We show that synthetic data, generated by Text2IMU, can be used to classify activities across three public benchmark datasets. We demonstrate that our Text2IMU synthesis approach does not require measured data of the target domain. Text2IMU yields an average Human Activity Recognition (HAR) accuracy of 79.2\% for correctly synthesised activities, which doubles the performance of synthetic sensor data obtained from baseline models. We demonstrate that synthetic HAR model training can replace empirical data acquisition when the prompted activities can be successfully generated.

Track: 3. Signal processing, machine learning, deep learning, and decision-support algorithms for digital and computational health

NominateReviewer: Lena Uhlenberg [e-mail]

Submission Number: 97