Publicação
Deep learning-based gait event detection andevaluationof gait-related spatiotemporal parameters in a robotic walker
| Resumo: | Gait disorders, especially those caused by neurological conditions such as cerebellar ataxia, can significantly affect a person’s mobility, independence, and overall quality of life. This dissertation introduces a deep learning-based gait event detection tool to accurately identify key real-time gait phases and transitions, supporting rehabilitation efforts. This research uses a hybrid deep learning model, named Conv3DL- STM, which combines the CNN and LSTM architectures. The model was trained and tested using RGB video recordings of individuals walking with WALKit Smart Walker in a fixed trajectory, achieving promising results—a test accuracy of 95.32%, precision of 93.16%, an f1-score of 93.14%, and a recall of 93.15%. These results reflect the system’s reliable ability to detect gait events in real-world conditions. This research also proposes a HIL assistive control strategy that uses an algorithm that estimates online and offline gait temporal parameters and dynamically adjusts the walker’s linear velocity in real time. By fine-tuning speed transitions based on real-time gait temporal parameter estimations—with of- fline estimation errors of 6.84% for step times, 3.62% for stride times, 5.21% for double support phase times, and 15.44% for cadence values, for unfiltered predictions—the proposed strategy creates a safer and more comfortable walking experience. This strategy is supported by the successful correlation be- 2 tween the user’s cadence and gait speed proven by a mean value of R of 0.83. In parallel, it provides multidisciplinary healthcare teams with precise data-driven insights, helping them monitor progress and personalize treatment plans. The system’s validation with an unknown healthy participant highlights the potential of deep learning to revolutionize gait training, with more personalized and user-specific rehabilitation while easing the workload of healthcare professionals. Future improvements will focus on expanding the dataset to include a wider range of gait impairments, refining the assistive control strategy for increased pathological flexibility, and conducting clinical trials to explore its long-term benefits for rehabilitation. |
|---|---|
| Autores principais: | Oliveira, Diogo Emanuel Nogueira de |
| Assunto: | Assistive control strategies Cerebellar ataxia Deep learning Gait event detection Gait rehabilitation Smart walker Deteção de eventos de marcha Aprendizagem profunda Andarilhos inteligentes Ataxia cerebela Reabilitação da marcha Estratégias de controlo assistido |
| Ano: | 2025 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade do Minho |
| Idioma: | inglês |
| Origem: | RepositóriUM - Universidade do Minho |
| Resumo: | Gait disorders, especially those caused by neurological conditions such as cerebellar ataxia, can significantly affect a person’s mobility, independence, and overall quality of life. This dissertation introduces a deep learning-based gait event detection tool to accurately identify key real-time gait phases and transitions, supporting rehabilitation efforts. This research uses a hybrid deep learning model, named Conv3DL- STM, which combines the CNN and LSTM architectures. The model was trained and tested using RGB video recordings of individuals walking with WALKit Smart Walker in a fixed trajectory, achieving promising results—a test accuracy of 95.32%, precision of 93.16%, an f1-score of 93.14%, and a recall of 93.15%. These results reflect the system’s reliable ability to detect gait events in real-world conditions. This research also proposes a HIL assistive control strategy that uses an algorithm that estimates online and offline gait temporal parameters and dynamically adjusts the walker’s linear velocity in real time. By fine-tuning speed transitions based on real-time gait temporal parameter estimations—with of- fline estimation errors of 6.84% for step times, 3.62% for stride times, 5.21% for double support phase times, and 15.44% for cadence values, for unfiltered predictions—the proposed strategy creates a safer and more comfortable walking experience. This strategy is supported by the successful correlation be- 2 tween the user’s cadence and gait speed proven by a mean value of R of 0.83. In parallel, it provides multidisciplinary healthcare teams with precise data-driven insights, helping them monitor progress and personalize treatment plans. The system’s validation with an unknown healthy participant highlights the potential of deep learning to revolutionize gait training, with more personalized and user-specific rehabilitation while easing the workload of healthcare professionals. Future improvements will focus on expanding the dataset to include a wider range of gait impairments, refining the assistive control strategy for increased pathological flexibility, and conducting clinical trials to explore its long-term benefits for rehabilitation. |
|---|