Development and Evaluation of Machine Learning Models for Human Body Model Generation - Machine Learning-Based Surface Positioning for Human Body Models - Master’s Thesis 2025

Aug 4, 2026

Upload by Ophelia

Finite element human body models (HBMs) are crucial numerical tools in biomechanics, medical simulation, and vehicle safety. Machine learning offers new ways to support finite element modeling, yet automated positioning of HBMs remains underexplored. This thesis presents data-driven skin-surface positioning driven by joint rotation angles, focusing on limited training data. Using only 144 samples, multiple models and smoothing strategies are implemented and evaluated through coordinate error, surface smoothness, geometric similarity, and local-shape preservation.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages49

Reading Time2 h 3 min

AccessFree to download

Abstract

Preface

Abbreviations

Introduction

Background

Research Question and Hypothesis

Methodology

Data Collection and Processing

Machine Learning Models

Model Refinement Strategies

Model Evaluation Metrics

Results

Performance of Basic Machine Learning Models

After Applying Refinement Strategies

Discussion

Modeling Challenges from Input-Output Imbalance

Basic Machine Learning Models Comparison

Smoothing Methods Comparison

Generalizability, Ethics, and Sustainability Considerations

Conclusion and Future Work

Summary of Contributions

Limitations

Future Work

Master’s Programme in ICT Innovation (EIT Digital Master School) Development and Evaluation of Machine Learning Models for Human Body Model Generation Machine Learning-Based Surface Positioning for Human Body Models Zili Zhou Master’s Thesis 2025 © 2025 This work is licensed under a Creative Commons“Attribution-NonCommercial-ShareAlike 4 .0 International” license. | Author Zili Zhou | | --- | | Title Development and Evaluation of Machine Learning Models for Human Body Model Generation—Machine Learning-Based Surface Positioning for Human Body Models | | Degree programme ICT Innovation (EIT Digital Master School) | | Major Autonomous Systems | | Supervisor Prof. Quan Zhou | | Advisors Prof. Haibo Li, Associate Prof. Xiaogai Li | | Date 10 June 2025 Number of pages 45+4 Language English | | Abstract
Finite element human body models (HBMs) are essential numerical tools in fields such as biomechanics, medical simulation, and vehicle safety. Recent developmentsin machine learning have opened up new opportunities in many directions relating to finite element modeling. However, its potential for automating the positioning of finite element HBMs remains largely unexplored. This project explores data-driven approaches for positioning the skin surface of HBMs based on joint rotation angle inputs, especially under limited training data conditions. This work complements efforts on full-body positioning by providing insights at the surface level for evaluating machine learning models-based positioning strategies, serving as a potential upstream step toward more comprehensive pipelines, such as the one in related full-body HBM work.
The dataset used for this project included only 144 samples, each involving changes in only one or two of 15 defined joints. To address this, this project implemented and evaluated multiple machine learning models, including Multi-Layer Perceptron (MLP), Random Forest, Conditional Variational Autoencoder (CVAE), Per Node MLP, and Graph Neural Network (GNN) . This project also explored methods for enhancing the skin surface smoothness of predicted HBMs, including incorporating Laplacian smoothing loss during training, applying post-processing filters (Laplacian, Taubin, and Bilateral), and using a combined MLP-GNN model. In this project, a comprehensive evaluation framework was also developed to assess the predicted skin surface of HBMs based on mean squared error (MSE) of coordinates, surface smoothness, geometric similarity, and preservation of the local shape of the model.
Experimental results showed that the combination of MLP with a GNN-based smoother achieved the best overall performance. Notably, despite training on a small dataset, the final model successfully positioned high-quality skin surfaces of HBMs for arbitrary joint configurations, which demonstrated the strong generalization capability of the proposed approach under limited data conditions. | | Keywords Machine Learning, Model Evaluation, Human Body Model, Shape Modeling, Surface Positioning, Smoothing Methods | Preface I would like to express my deepest gratitude to my supervisors, Siyuan Chen and Xiaogai Li from KTH Royal Institute of Technology. They provided the complete dataset for this project and offered patient, insightful guidance throughout the process. Without their support and expertise, this thesis would not have been possible. I am also sincerely thankful to the two universities I attended during my master’s studies: KTH Royal Institute of Technology and Aalto University. Both institutions offered me a rich academic environment and invaluable learning experiences. This journey has been one of the most memorable and formative periods of my life. Lastly, I would like to thank all the teachers, friends, and especially my family, who have supported me along the way. Their encouragement has always been a constant source of motivation and strength throughout this master program. Contents Abstract 3 Pref

FAQs

What problem does the thesis address in human body model generation?

It addresses data-driven automation for positioning the skin surface of finite element human body models using joint rotation angles. The focus is on achieving reliable results even when training data are limited.

Which machine learning models are implemented and evaluated?

The thesis evaluates several models including MLP, Random Forest, Conditional Variational Autoencoder (CVAE), Per Node MLP, and Graph Neural Network (GNN). It also studies a combined MLP-GNN approach for smoothing.

How is the quality of the predicted skin surface assessed?

Evaluation uses multiple metrics, including mean squared error of coordinates, surface smoothness, geometric similarity, and preservation of local shape of the model.

Aug 4, 2026

Upload by Ophelia