Gait data generation using lightweight generative deep learning framework.

Area of Science:

• Biomechanics
• Machine Learning
• Robotics

Background:

• Human gait analysis is crucial for sports science, exoskeleton design, and clinical applications.
• Collecting gait data is challenging due to physiological and ethical constraints, leading to data scarcity.
• Existing deep learning models, like Generative Adversarial Networks (GANs), are often computationally intensive and impractical for real-world use.

Purpose of the Study:

• To develop a novel, lightweight hybrid model for efficient human gait data generation.
• To address the limitations of computational intensity and data scarcity in current gait analysis methods.
• To create a model that balances complexity and data fidelity for practical applications.

Main Methods:

• Proposed a novel hybrid model combining a Feed-forward Neural Network (FNN) and an Autoencoder (AE), termed FNN-AE.
• The FNN component generates initial gait data, while the AE refines it for enhanced realism.
• Verified the model's physical plausibility using Newtonian equations of motion and biomechanical simulations on the OpenSim platform.

Main Results:

• The FNN-AE model achieves satisfactory performance comparable to state-of-the-art methods.
• The proposed architecture significantly reduces model complexity by utilizing fewer parameters.
• Generated gait data demonstrated biomechanical feasibility when tested in simulations.

Conclusions:

• The FNN-AE model presents an effective and computationally efficient solution for human gait data generation.
• This lightweight approach overcomes the practical limitations of existing deep learning models for gait analysis.
• The model shows promise for advancing research and applications in sports science, robotics, and clinical settings.