AdversarialGait: direction invariant gait recognition with adversarial learning

Hilton Raj, A. V. Vishnuram, Rahul Raman

Abstract: Gait recognition is an emerging biometric technology that identifies individuals based on their unique walking patterns with gait sequences and variations serving as strong biometric features. Unlike other biometric modalities, it can operate over long distances without requiring active participation from the subjects, thus having wide application in security and surveillance. The performance of gait recognition can be significantly affected by variations such as view angle, posture, clothing and occlusion. Despite the advances in deep learning, these variations still pose a challenge. Specifically, a person’s appearance differs based on walking directions, impacting the accuracy of gait recognition. Existing works primarily address appearance-level variation and do not explicitly suppress the influence of walking direction. We propose a novel approach to remove the influence of walking direction in gait recognition via an adversarial process. We introduce a three module framework—the Feature Extraction Network (FEN), Gait Recognition Network (GRN) and the Direction Estimation Network (DEN). We implement an adversarial training paradigm with walking direction as the adversarial parameter in which we simultaneously train the FEN to enhance recognition capabilities of GRN while hindering DEN’s ability to estimate the walking direction, thus learning direction irrelevant gait recognition features. We train our model on the benchmark datasets, CASIA-B and OU-MVLP. Furthermore, we provide the first experimental demonstration showing that adversarial learning actively avoids direction-related features forming more compact clusters with better inter-class separability while maintaining comparable accuracy to state-of-the-art gait recognition models.

External IDs:dblp:journals/nca/RajVR25