Gait Trajectory Prediction on an Embedded Microcontroller Using Deep Learning.

Area of Science:

• Robotics and Artificial Intelligence
• Biomedical Engineering
• Embedded Systems

Background:

• Active prosthesis control is complex due to gait trajectory kinematics.
• Artificial Neural Networks (ANNs) are effective for gait trajectory generation but computationally intensive.
• Deploying complex ANNs on embedded systems requires mobility and reliability solutions.

Purpose of the Study:

• Investigate simpler, reliable ANNs for active prosthesis control on a low-cost microcontroller.
• Evaluate Multi-Layered Perceptron (MLP) and Convolutional Neural Network (CNN) performance for gait trajectory prediction.
• Enable system mobility by deploying ANN algorithms on a single embedded device.

Main Methods:

• Trained MLP and CNN models on shank and foot Inertial Measurement Unit (IMU) data from four subjects.
• Tested models on a fifth subject to predict gait trajectory 200 ms ahead, with and without current gait phase information.
• Deployed trained models on a low-cost ESP32 microcontroller for real-time performance evaluation.

Main Results:

• CNN achieved a higher correlation coefficient (0.973) than MLP (0.945) with limited data (excluding gait phase).
• Both MLP and CNN achieved high correlation coefficients (approx. 0.98) when current gait phase was provided.
• MLP execution time (2.4 ms) was significantly faster than CNN (142 ms) on the microcontroller.

Conclusions:

• CNN is more efficient for gait prediction with scarce data, offering acceptable execution times.
• MLP provides relative accuracy with low execution time when sufficient training data is available.
• The study demonstrates the feasibility of deploying simplified ANNs on microcontrollers for mobile active prosthesis control.