MoCap to IMU Knowledge Distillation for Activity Recognition

MSc Thesis Proposal

Proposed Analysis Tools

Alba Huti

LARa and OpenPack Datasets

Synchronized MoCap plus IMU recordings

MS-G3D Teacher Model

Spatio-temporal graph convolution on MoCap skeleton

Lightweight TCN / LSTM Student

Compact 2-4 IMU edge model

Knowledge Distillation

KL-divergence plus MSE multi-objective loss

Robustness Evaluation

Cross-user, cross-scenario, sensor perturbation stress tests

Deployment Metrics

Latency, model size, macro-F1, RMSE

MSc Thesis Proposal

Calendar of Activities

Gantt Chart

Alba Huti

1. Literature Review and Dataset Preparation

2. Teacher Model (MS-G3D) Implementation

3. Student Model (TCN/LSTM) Design

4. Knowledge Distillation Training

5. Robustness Evaluation and Stress Testing

6. Deployment Feasibility Testing

7. Thesis Writing and Finalization

MSc Thesis Proposal

Expected Contribution

Alba Huti

1

Reproducible MoCap-to-IMU Transfer Pipeline

Concrete teacher-student distillation framework for temporal motion data with defined training procedures.

2

Robustness Evidence

Empirical proof that MoCap supervision improves cross-user and cross-scenario generalization beyond standard IMU-only training.

3

Minimal Sensor Configuration Guidance

Sensor count vs. performance trade-off curves identifying the minimal sufficient IMU setup for industrial deployment.

4

Standardized Benchmarking Protocol

Cross-user/cross-scenario splits, synthetic perturbations, and metrics (macro-F1, RMSE, latency, model size) as a reference framework.

5

Deployment-Oriented Insights

Quantitative linking of sensor count, robustness, inference latency and memory footprint for embedded exoskeleton systems.