Resumo:
This work addresses the development of optimized Convolutional Neural Network (CNN)
models for knee angle prediction in embedded systems, targeting applications in lower
limb prosthetics. By leveraging low-cost Inertial Measurement Units (IMUs) and multiobjective
optimization techniques, specialized models were developed for different gait
types (short, natural, and long strides), associated with a classifier capable of real-time
gait pattern identification. The results demonstrate that specialized models significantly
outperform a combined model, achieving a root mean square error (RMSE) of 2.050◦,
an improvement of over 48% compared to the latter. Furthermore, implementation on
a Kendryte K210 microcontroller validated the feasibility of deployment on accessible
hardware while maintaining the computational efficiency required for real-time applications.
This study contributes to the advancement of assistive technologies, indicating that efficient
deep learning models can be effectively implemented on embedded systems, improving
the quality of life for individuals with lower limb amputations. This approach further
increases the affordability of advanced prosthetic solutions, making them more accessible
to individuals in economically disadvantaged regions.
