Publicação
Real-time patient pose estimation on a smart walker
| Resumo: | Rehabilitation is important to improve quality of life for mobility impaired patients. Smart walkers are commonly used to provide residual motor skills recovery, based on repetitive and intensity-adapted training. These should embed automatic, objective and real-time tools for user-centered control and monitoring. Yet, present solutions have focused only on extracting few very specific metrics using dedicated sensors with no unified full-body approach. This dissertation proposes the creation of a general, real-time and robust full-body spatial pose estimation solution using visual information from two camera streams, with non-overlapping ROIs, mounted on the ASBGo smart walker used in patient rehabilitation. Human joint estimation is performed using a two-stage Neural Network framework, where keypoints are first detected in 2D image frames of both cameras, using a Fully Convolutional Network, and then lifted to 3D space relative to the walker, using a Fully Connected regression module. A custom acquisition method was also developed and used to obtain a dataset containing data from 14 healthy subjects, used for training and evaluating the proposed solution offline, which was then deployed and integrated on the real smart walker. An overall detection error of 3.73 pixels and 44.05mm were reported for each stage respectively, with an inference time of 26.6ms when deployed on the constrained hardware of the equipment, during normal use. The final solution was able to extract a compact body representation from inexpensive sensors, which can be used as a common base to calculate full patient gait and posture metrics, allow Human-Robot interaction applications and human-in-the loop control for personalized rehabilitation. Despite promising results, more data should be collected with impaired subjects, in order to assess the model’s true performance as a rehabilitation tool in real-world scenarios. |
|---|---|
| Autores principais: | Palermo, Manuel de Castro |
| Assunto: | Rehabilitation Smart walker Human pose estimation Computer vision Machine learning Reablitação Andarilho inteligente Deteção da postura humana Visão computador Aprendizagem máquina |
| Ano: | 2021 |
| 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: | Rehabilitation is important to improve quality of life for mobility impaired patients. Smart walkers are commonly used to provide residual motor skills recovery, based on repetitive and intensity-adapted training. These should embed automatic, objective and real-time tools for user-centered control and monitoring. Yet, present solutions have focused only on extracting few very specific metrics using dedicated sensors with no unified full-body approach. This dissertation proposes the creation of a general, real-time and robust full-body spatial pose estimation solution using visual information from two camera streams, with non-overlapping ROIs, mounted on the ASBGo smart walker used in patient rehabilitation. Human joint estimation is performed using a two-stage Neural Network framework, where keypoints are first detected in 2D image frames of both cameras, using a Fully Convolutional Network, and then lifted to 3D space relative to the walker, using a Fully Connected regression module. A custom acquisition method was also developed and used to obtain a dataset containing data from 14 healthy subjects, used for training and evaluating the proposed solution offline, which was then deployed and integrated on the real smart walker. An overall detection error of 3.73 pixels and 44.05mm were reported for each stage respectively, with an inference time of 26.6ms when deployed on the constrained hardware of the equipment, during normal use. The final solution was able to extract a compact body representation from inexpensive sensors, which can be used as a common base to calculate full patient gait and posture metrics, allow Human-Robot interaction applications and human-in-the loop control for personalized rehabilitation. Despite promising results, more data should be collected with impaired subjects, in order to assess the model’s true performance as a rehabilitation tool in real-world scenarios. |
|---|