arrow
Return

Multiple motion pattern augmentation assisted gait recognition

delete2025-07-05
delete0
PRE
AI
W
Wei Huo
J
Jun Tang
W
Wenxia Bao
K
Ke Wang *
N
Nian Wang
L
Liang Dong
DOI:10.1016/j.sigpro.2025.110185delete
deleteOriginal
deleteOriginal request for help
deleteShare
deleteSave
Abstract

Abstract

En 中文
• We present a novel motion pattern augmentation module MPA to generate diverse gait activities based on the raw silhouettes. Different from the existing data augmentation methods in gait recognition, we focus on constructing some novel motion sequences containing distinctive body movement habits instead of applying simple changes at the image-level. Moreover, in MPA, we first separately extract motion features from various augmented gait sequences, and then compress them into a compact fusion feature, which helps the model learn the discriminative identity representations from the multiple augmented motion sequences. • We propose a novel gait recognition framework named GaitMPA, which is mainly composed of MFNet, MPA, and MFA. MFNet is designed to capture dynamic motion difference and learn the multi-grained spatio-temporal representations of different body parts. Moreover, based on the augmented motion sequences generated by MPA, we present MFA to merge the multi-source features of raw gait sequence and augmented motion data, and aggregate them into a more powerful gait feature. Finally, the outputs of MFNet and MFA are fused for the final identity recognition. • We conduct extensive experiments on five public datasets, i.e., CASIA-B, OU-MVLP, CCPG, GREW, and Gait3D. The results of comparative experiments and ablation studies demonstrate the effectiveness and superiority of the proposed method.

Journal

Signal Processing cover
Signal Processing
IF:
3.6
Papers:
9.9K
Citations:
1.7W

Organization

A
anhui university
Scholars:
1.9W
Papers: 1.2W
Citations: 24