Approach to a Lower Body Gait Generation Model Using a Deep Convolutional Generative Adversarial Network

Research over gait analysis has become more relevant in the last years, especially as a tool to detect early frailty signs. However, data gathering is often difficult and requires lots of resources. Synthetic data generation is a great complementary tool for data gathering that enables the augmentation of existing datasets. Despite not being a new concept, it has gained popularity in the last years thanks to Generative Adversarial Networks (GANs), a neural network architecture capable of creating data indistinguishable from the original one. In this article deep-convolutional GANs has been used to artificially expand a gait dataset containing data of the lower part of the body. The synthetic data has been studied through three approaches: looking animations of the points and comparing them to the originals; applying principal component analysis algorithm to both datasets to visually assess how each of them is distributed; and by extracting different features from both datasets to compare their statistical differences. The evaluation showed promising results, which opens a path for using synthetic data generation in the gait analysis domain.