Background
Human Activity Recognition (HAR) using wearable IMUs is widely used in healthcare, sports, and ubiquitous computing. However, collecting labeled IMU datasets is costly and limits scalability and generalization.
Recent advances in two areas open up new possibilities:
Modern generative models can synthesize realistic human motion sequences from natural language descriptions.
While motion generation and IMU simulation are well studied independently, there is little work combining language-driven motion generation with synthetic IMU data for HAR.
This thesis explores whether natural language can serve as a control interface for generating semantically meaningful training data.
Your Tasks
1. Literature Review
-
Study text-to-motion models and IMU simulation methods
-
Identify suitable tools and datasets
2. Motion Generation from Text
-
Generate motion sequences using a text-to-motion model
-
Design prompts for different activities and variations
3. IMU Signal Simulation
-
Convert motion trajectories into:
-
accelerometer data
-
gyroscope data
-
-
Incorporate realistic factors:
-
sensor placement
-
noise and variability
-
4. Synthetic Dataset Creation
-
Generate datasets using different prompt strategies:
-
fixed prompts
-
diverse/augmented prompts
-
5. Model Training
-
Train activity recognition models using:
-
real data
-
synthetic data
-
hybrid combinations
-
6. Evaluation
-
Evaluate on standard datasets (e.g., UCI HAR, PAMAP2)
-
Analyze:
-
performance on real data
-
generalization to unseen activities
-
data efficiency (few-shot scenarios)
-
Requirements
-
Python programming
-
Basic machine learning / deep learning knowledge
-
Some familiarity with time-series data processing
Application Documents
- A paragraph explaining your motivation.
- Your study program (Bachelor/Master), current semester, and field of study.
- A transcript of records (courses and grades).
- Your programming experience.
- Any areas of interest relevant to the topic.
- Your CV (if available)
