کنفرانس مهندسی برق ایران

صفحه اصلی / سی و سومین کنفرانس بین المللی مهندسی برق

Real-Time Prediction of Lower Limb AngularTrajectories Using an Optimized LSTM Model withMarkerless Motion Capture

نویسندگان :

Amirhossein Jafari¹ Hamed Jalaly Bidgoly²

1- دانشگاه صنعتی اصفهان 2- دانشگاه صنعتی اصفهان

کلمات کلیدی :

Gait analysis،Joint angle prediction،Markerless motion capture،Long Short-Term Memory (LSTM)،Assistive devices

چکیده :

Lower limb dysfunction affects millions of people worldwide and often requires the use of assistive devices to restore mobility. While passive devices provide support, they lack the adaptability of active devices, which utilize actuated joints for more natural locomotion. The active devices rely on control architectures that predict angular gait trajectories, ensuring precise synchronization with human motion. This study proposes a Long Short-Term Memory (LSTM)-based model to predict the angular trajectories of the ankle, knee, and thigh during walking. The model was trained using data from a markerless motion capture system, eliminating the need for wearable sensors and simplifying the data acquisition process. Optimization techniques such as pruning and quantization were applied to enhance real-time performance in resource-constrained assistive devices. Additionally, we developed an algorithm to address inaccuracies in the pose estimator’s joint angle predictions caused by challenging capturing conditions. This algorithm improves data reliability through outlier detection and correction. The experimental results demonstrate that the optimized model accurately predicts gait trajectories across varying walking speeds, showcasing its potential for integrating active assistive devices to improve mobility and adaptability.Lower limb dysfunction affects millions of people worldwide and often requires the use of assistive devices to restore mobility. While passive devices provide support, they lack the adaptability of active devices, which utilize actuated joints for more natural locomotion. The active devices rely on control architectures that predict angular gait trajectories, ensuring precise synchronization with human motion. This study proposes a Long Short-Term Memory (LSTM)-based model to predict the angular trajectories of the ankle, knee, and thigh during walking. The model was trained using data from a markerless motion capture system, eliminating the need for wearable sensors and simplifying the data acquisition process. Optimization techniques such as pruning and quantization were applied to enhance real-time performance in resource-constrained assistive devices. Additionally, we developed an algorithm to address inaccuracies in the pose estimator’s joint angle predictions caused by challenging capturing conditions. This algorithm improves data reliability through outlier detection and correction. The experimental results demonstrate that the optimized model accurately predicts gait trajectories across varying walking speeds, showcasing its potential for integrating active assistive devices to improve mobility and adaptability.