SSGait: enhancing gait recognition via semi-supervised self-supervised learning

Gait recognition is a challenging biometric technology field due to the complexity of integrating static appearance and dynamic movement patterns in walking videos and the need for extensive labeled data. To address these challenges, we propose an effective self-supervised semi-supervised gait recog...

Full description

Saved in:
Bibliographic Details
Published in:Applied intelligence (Dordrecht, Netherlands) Netherlands), 2024-04, Vol.54 (7), p.5639-5657
Main Authors: Xi, Hao, Ren, Kai, Lu, Peng, Li, Yongqiang, Hu, Chuanping
Format: Article
Language:English
Subjects:
Algorithms
Artificial Intelligence
Computer Science
Crime prevention
Gait recognition
Learning
Machine learning
Machines
Manufacturing
Mechanical Engineering
Methods
Morphology
Processes
Representations
Self-supervised learning
Semantics
Surveillance
Walking
Citations: Items that this one cites
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Gait recognition is a challenging biometric technology field due to the complexity of integrating static appearance and dynamic movement patterns in walking videos and the need for extensive labeled data. To address these challenges, we propose an effective self-supervised semi-supervised gait recognition (SSGait) method for learning spatiotemporal representations. Specifically, SSGait contains two main branches: the semi-supervised branch varies the edge morphology of the gait silhouette by introducing sequence-level morphological perturbations in the input, and it learns robust representations using the mean teacher architecture. In the self-supervised branch, we devise two pretexts closely aligned with the gait recognition task: tube-masked reconstruction and clip order prediction. In this way, SSGait can improve the accuracy of recognition tasks by utilizing features extracted from a substantial volume of unlabeled gait data. Finally, we extensively evaluate the proposed SSGait algorithm on two widely used cross-view gait datasets, namely, CASIA-B and OU-MVLP. The experimental results show that SSGait outperforms the best fully supervised methods by 4.84% when using only 20% of the labeled data. When using only 80% of the labeled data, SSGait does match the performance of fully supervised methods trained with 100% labeled data. This highlights SSGait’s ability to use fewer labeled data effectively for improved gait recognition.
ISSN:0924-669X
1573-7497
DOI:10.1007/s10489-024-05385-2