Towards a generic deep model: Human gait analysis using wearable devices for smart health care in internet of health things

Abstract

Human gait is a fascinating subject due to its unique bipedal nature, which encodes valuable information and patterns within its sparse kinematics. Previous research has demonstrated that low-level signals captured by wearable devices equipped with inertial measurement units can estimate soft biometrics, detect emotional states, classify activ ities of daily living, and facilitate person re-identification using machine/deep learning models. However, no prior efforts have been made to develop a generic deep-learning model that can analyze human gait signatures and estimate these various aspects us ing wearable devices. In this study, we propose a novel RNN-CNN neural network inspired by ResNeXt and UltaNet that is capable of analyzing human gait signatures and estimating these various aspects using wearable devices. Our model is generic in nature, and we have trained it using several Human Activity Recognition datasets, including WISDM 2011, WISDM Actitracker, and WISDM 2019. We achieved accura cies of 95.624%, 96.978%, and 82.415%, respectively, on these datasets. Furthermore, we trained our model on a locally generated dataset for emotion recognition, where it achieved an accuracy of 78.198%. We also conducted experiments on another locally generated dataset for person Re-identification, achieving an accuracy of 93.713%. These results demonstrate the effectiveness of our proposed RNN-CNN model for analyzing human gait signatures and estimating various aspects using wearable devices.