Generative Data Augmentation of Human Biomechanics

Wearable sensors are miniature and affordable devices used for monitoring human motion in daily life. Data-driven models applied to wearable sensor data can enhance the accuracy of movement analysis outside of controlled settings. However, obtaining a large and representative database for training these models is challenging due to the specialised motion laboratories and expensive equipment required. To address this limitation, this study proposes a data augmentation approach using generative deep learning to enhance biomechanical datasets. A novel conditional generative adversarial network (GAN) was developed to synthesise biomechanical data during gait. The GAN takes into account the subject’s anthropometric measures to generate data that represents specific body types as well as information about the gait cycle for reconstruction back into the time domain. The proposed model was evaluated for generating biomechanical data of unseen subjects and fine-tuning the model with small percentages (1%, 2% and 5%) of the test dataset. Researchers and practitioners can overcome the limitations of obtaining large training datasets from human participants by synthesising realistic and diverse synthetic data. This paper outlines the methodology and experimental setup for developing and evaluating the GAN and discusses its potential impact on the field of biomechanics and human motion analysis.