Residual Graph Convolutional Neural Network for Gait Recognition across Various Walking Conditions

Abstract

Over the years, extensive attention was given to a person identification task to prevent fraudulent activities. Several techniques have been developed for verifying persons from video or image by considering biometrics such as face, palm print, iris, and gait. Among all the traits mentioned above, gait is an unobtrusive and easily collectible biometric that can be observed without hindrance to the subject's activity. However, gait recognition performance can deteriorate under challenging conditions, including unconstrained path, bulky clothing, and different viewing angles. To provide an effective solution to gait recognition under these conditions, this thesis pioneers developing the Residual Connection-based Graph Convolutional Neural Network architecture for robust and reliable gait recognition. The proposed methodology incorporates residual connections for gait recognition from videos. Furthermore, the proposed system is lightweight in terms of computational cost, making the model deployable in practice. CASIA-B and AVA multi-view Gait datasets are used to evaluate the efficacy of the proposed method. The developed system attained 97.03% mean accuracy under normal walking conditions, 90.77% mean accuracy under bag carrying conditions, 89.90% mean accuracy under bulky clothes wearing conditions on CASIA-B gait dataset, and 98.85% mean accuracy for unconstrained gaits on AVA multi-view Gait dataset. The findings demonstrate that the proposed methodology outperformed other state-of-the-art gait recognition systems under challenging walking conditions.