A Novel Approach to Abnormal Gait Recognition Based on Generative Adversarial Networks

Abstract

It is crucial to accurately recognize abnormal gaits of users with weak motion capability during the process of interacting with a robot. However, the method of recognizing abnormal gaits through wearable sensors has limitations when the elderly and disabled are assisted. Meanwhile, the non-contact recognition method of abnormal gaits based on deep learning requires a large number of labeled samples to solve underfitting and overfitting problems. In this paper, we propose a novel approach based on the combination of generative adversarial networks and deep convolutional neural networks to recognize abnormal gaits. Firstly, to obtain rich and varying abnormal gait images, we propose AGR-GAN, which classifies generated gait images in a supervised way and learns interpretable representations between latent variables and abnormal gait images in an unsupervised way. Secondly, we screen out the latent variables related to high-quality generated gait images. We input the latent variables and category labels into the trained AGR-GAN to obtain gait images of specific abnormal gait types with diverse postures and varying perspectives. Finally, we use the transfer-learning AGR-GAN discriminator as a gait recognition network to recognize multiple abnormal gait images. Through verification and comparison of real gaits, we obtain an accuracy rate of 88.9%. Therefore, our proposed abnormal gait recognition method based on generative adversarial networks increases the gait types and scale of the dataset, improving the accuracy of multiple gait recognition.