Covariate factor mitigation techniques for robust gait recognition

Abstract

The human gait is a discriminative feature capable of recognising a person by their unique walking manner. Currently gait recognition is based on videos captured in a controlled environment. These videos contain challenges, termed covariate factors, which affect the natural appearance and motion of gait, e.g. carrying a bag, clothing, shoe type and time. However gait recognition has yet to achieve robustness to these covariate factors. To achieve enhanced robustness capabilities, it is essential to address the existing gait recognition limitations. Specifically, this thesis develops an understanding of how covariate factors behave while a person is in motion and the impact covariate factors have on the natural appearance and motion of gait. Enhanced robustness is achieved by producing a combination of novel gait representations and novel covariate factor detection and removal procedures. Having addressed the limitations regarding covariate factors, this thesis achieves the goal of robust gait recognition. Using a skeleton representation of the human figure, the Skeleton Variance Image condenses a skeleton sequence into a single compact 2D gait representation to express the natural gait motion. In addition, a covariate factor detection and removal module is used to maximise the mitigation of covariate factor effects. By establishing the average pixel distribution within training (covariate factor free) representations, a comparison against test (covariate factor) representations achieves effective covariate factor detection. The corresponding difference can effectively remove covariate factors which occur at the boundary of, and hidden within, the human figure.