Gait recognition with covariate separation based on adversarial learning

Abstract

In response to the poor recognition accuracy caused by viewpoint changes and wearing changes in gait recognition problems, a viewpoint classifier and a wearing classifier are proposed to be connected after the feature extractor. The parameters of the feature extractor are updated using the gradient reversal mechanism so that its function is opposite to that of the two classifiers, thereby separating the viewpoint information, wearing information at the feature extraction level, and reducing the impact of these two most common factors on gait recognition. Comparative experiments on the CASIA-B dataset show that under BG and CL conditions, our model has improved by an average of 2.1% and 1.9% in cross-viewpoint accuracy compared to the baseline model. Our model is more robust under complex walking conditions.