Spatiotemporal smoothing aggregation enhanced multi-scale residual deep graph convolutional networks for skeleton-based gait recognition

Gait recognition has a variety of development potentials, such as noncontact potential. The preference for skeleton-based recognition arises due to challenges posed by self-occlusion and environmental factors affecting silhouette-based methods. Addressing the discriminative properties of long-term and short-term temporal cues, we propose spatiotemporal smoothing aggregation enhanced multiscale residual deep graph convolutional networks. This paper considers both long and short gait feature time series, enabling the learning of discriminative multiscale representations. In the baseline network, three scale features are sequentially extracted, followed by a reverse process to extract and fuse multiscale features. This method significantly bolsters the ability of graph convolution to effectively model the context knowledge of human poses effectively. This study investigated multiscale gait feature aggregation, which significantly mitigates oversmoothing effects. A spatiotemporal smoothing aggregation module with an embedded attention mechanism is introduced to hierarchically aggregate and enhance multiscale key joint features. This module alleviates oversmoothing in deep graph convolutional networks. The method underwent rigorous testing on the Chinese Academy of Sciences Institute of Automation(CASIA-B) dataset, achieving an average accuracy of 78.2%, ranking as the second highest performing skeletal-based gait recognition model currently available, and attaining rank-1 accuracies of 14.7 and 8.19 on Gait Recognition in the wild (GREW) and Gait3D datasets, respectively.