Development and Evaluation of a Machine-Learning Based Fall Detection System for Prosthetic Knees
(2024)Department of Automatic Control
- Abstract
- This thesis explores the feasibility of integrating a fall detection system into microprocessor-controlled prosthetic knees using onboard sensors, with a focus on optimizing machine learning models for real-time operational efficiency within the limited computational capacities of such devices. Initial investigations utilized the public UMAFall dataset to gain insights into fall detection methodologies and preprocessing techniques. This study also examined the potential for combining the UMAFall dataset with device-specific data to enhance model robustness and performance.
Several machine learning models, including Support Vector Machines (SVM), Logistic Regression (LR), and Random Forests (RF), were evaluated for their ability to... (More) - This thesis explores the feasibility of integrating a fall detection system into microprocessor-controlled prosthetic knees using onboard sensors, with a focus on optimizing machine learning models for real-time operational efficiency within the limited computational capacities of such devices. Initial investigations utilized the public UMAFall dataset to gain insights into fall detection methodologies and preprocessing techniques. This study also examined the potential for combining the UMAFall dataset with device-specific data to enhance model robustness and performance.
Several machine learning models, including Support Vector Machines (SVM), Logistic Regression (LR), and Random Forests (RF), were evaluated for their ability to accurately detect falls and for their suitability in terms of computational footprint when deployed in a prosthetic device environment. The models were initially trained with 42 features, which increased to 56 after incorporating pitch and roll estimations into the device-specific dataset. This study further experimented with reducing the feature set to 10 core features to examine the impact on model size and efficiency.
Results indicate that feature reduction significantly decreases model size while maintaining high accuracy, with SVM and LR models showing the most substantial reduction in size, making them ideal candidates for on-device implementation. The Random Forest model, although effective in fall detection, demonstrated a less significant reduction in size, posing challenges for its practical deployment in prosthetic knees with strict hardware limitations. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9173631
- author
- Magnúsdóttir, Heiðrún Dís
- supervisor
- organization
- year
- 2024
- type
- H3 - Professional qualifications (4 Years - )
- subject
- report number
- TFRT-6236
- other publication id
- 0280-5316
- language
- English
- id
- 9173631
- date added to LUP
- 2024-09-09 09:19:02
- date last changed
- 2024-09-09 09:19:02
@misc{9173631, abstract = {{This thesis explores the feasibility of integrating a fall detection system into microprocessor-controlled prosthetic knees using onboard sensors, with a focus on optimizing machine learning models for real-time operational efficiency within the limited computational capacities of such devices. Initial investigations utilized the public UMAFall dataset to gain insights into fall detection methodologies and preprocessing techniques. This study also examined the potential for combining the UMAFall dataset with device-specific data to enhance model robustness and performance. Several machine learning models, including Support Vector Machines (SVM), Logistic Regression (LR), and Random Forests (RF), were evaluated for their ability to accurately detect falls and for their suitability in terms of computational footprint when deployed in a prosthetic device environment. The models were initially trained with 42 features, which increased to 56 after incorporating pitch and roll estimations into the device-specific dataset. This study further experimented with reducing the feature set to 10 core features to examine the impact on model size and efficiency. Results indicate that feature reduction significantly decreases model size while maintaining high accuracy, with SVM and LR models showing the most substantial reduction in size, making them ideal candidates for on-device implementation. The Random Forest model, although effective in fall detection, demonstrated a less significant reduction in size, posing challenges for its practical deployment in prosthetic knees with strict hardware limitations.}}, author = {{Magnúsdóttir, Heiðrún Dís}}, language = {{eng}}, note = {{Student Paper}}, title = {{Development and Evaluation of a Machine-Learning Based Fall Detection System for Prosthetic Knees}}, year = {{2024}}, }