Graph convolutional networks (GCNs) are extensively used for skeleton‐based gait recognition. Nevertheless, despite significant improvements, a substantial challenge lies in the restricted receptive field of GCNs. However, separate structural joints could also reveal a notably important correlation. Prior research rarely investigates joints’ local and global relationships, resulting in inadequate modeling of the complex dynamics of skeletal sequences. To address this issue, we propose a GCN and self‐attention dynamic fusion network (GSDFN), GSDF‐Gait, which combines the GCN with a Self‐attention (SA) mechanism in parallel to address the challenge of modeling long‐range skeleton joint correlations in gait recognition. The graph SA comprises paired SA, which presents the interrelationship between each pair of bodily joints. A multivariate spatial and temporal feature (MVSTF) approach is introduced to extract geometrical and directional features from spatial and temporal dimensions. The temporal convolution network (TCN) investigates the relationships among temporal joint frames. The spatial and temporal model covers the joints’ temporal behaviors and long‐range dependency. The model is evaluated based on the CASIA‐B, OUMVLP‐Pose, and GREW datasets. On the CASIA‐B dataset, we achieved significant accuracy of 97.10%, 93.20%, and 90.80% on normal walking, carrying bags, and wearing clothes, respectively, whereas our model achieved 92.90% and 73.5% on the OUMVLP‐Pose and GREW datasets, respectively.
Khaliluzzaman et al. (Thu,) studied this question.