Gait data generation using lightweight generative deep learning framework

Human gait analysis is an intriguing area of research that supports sports science, robotic exoskeleton design, and clinical applications. However, the collection of gait data is a challenging task under physiological and ethical conditions, which leads to data scarcity. Many clinical works have utilized gait data generation using deep learning frameworks like Generative Adversarial Network (GAN) to address these limitations. However, most of these models are computationally intensive, which makes them impractical for real-world scenarios. To address these challenges, we propose a novel lightweight hybrid model ensembling a Feed-forward Neural Network with an Autoencoder, termed as FNN-AE model. The proposed architecture is designed to balance between model complexity and data fidelity. The FNN generates the gait data, while the AE refines it to resemble the real gait patterns closely. This model achieves satisfactory performance with state-of-the-art models while utilizing fewer parameters to reduce complexity. The proposed model is verified with the Newtonian equation of motion. The model generated data are tested on the OpenSim simulation platform to check the biomechanical feasibility of generated gait patterns.