AggreGait: Automatic gait feature extraction for human age and gender classification with possible occlusion

Abstract

The growing interest in smart surveillance and automated public access control necessitates robust age and gender classification (AGC) techniques that can operate effectively in unconstrained environments. While model-based gait obtained via pose estimation offers a promising approach, its performance can be hindered by occlusions commonly encountered in real-world videos. In this work, we propose a custom Graph Neural Network (GNN) architecture, AggreGait, for robust AGC under occlusions. AggreGait integrates upper and lower body features with whole-body information for age and gender prediction. We train AggreGait on pose sequences from the gait-in-the-wild (GITW) dataset, simulating different types of occlusions. AggreGait performs comparably to existing methods, achieving an overall accuracy of 91% in unobstructed conditions. Notably, AggreGait maintains reasonable accuracy using only upper limb (or upper and lower limb) features, suggesting its potential for real-time surveillance applications despite occlusions. This work paves the way for practical gait-based AGC in unconstrained environments, enhancing the effectiveness of surveillance systems and facilitating automated access control.