This website requires JavaScript.

Human Health Indicator Prediction from Gait Video

Ziqing LiXuexin YuXiaocong LianYifeng WangXiangyang Ji
Dec 2022
Body Mass Index (BMI), age, height and weight are important indicators ofhuman health conditions, which can provide useful information for plenty ofpractical purposes, such as health care, monitoring and re-identification. Mostexisting methods of health indicator prediction mainly use front-view body orface images. These inputs are hard to be obtained in daily life and often leadto the lack of robustness for the models, considering their strict requirementson view and pose. In this paper, we propose to employ gait videos to predicthealth indicators, which are more prevalent in surveillance and home monitoringscenarios. However, the study of health indicator prediction from gait videosusing deep learning was hindered due to the small amount of open-sourced data.To address this issue, we analyse the similarity and relationship between poseestimation and health indicator prediction tasks, and then propose a paradigmenabling deep learning for small health indicator datasets by pre-training onthe pose estimation task. Furthermore, to better suit the health indicatorprediction task, we bring forward Global-Local Aware aNd CentrosymmetricEncoder (GLANCE) module. It first extracts local and global features byprogressive convolutions and then fuses multi-level features by acentrosymmetric double-path hourglass structure in two different ways. Experiments demonstrate that the proposed paradigm achieves state-of-the-artresults for predicting health indicators on MoVi, and that the GLANCE module isalso beneficial for pose estimation on 3DPW.