Jeongho Kim, Mingyu Lee, Jinwook Kim, Joonho Seon, Jin Young Kim
언어
영어(ENG)
URL
https://www.earticle.net/Article/A481189
※ 원문제공기관과의 협약기간이 종료되어 열람이 제한될 수 있습니다.
원문정보
초록
한국어
Recently, various deep learning-based approaches have been studied to interpret complex human actions. A significant limitation of conventional methods is their reliance on a fixed graph structure, which restricts their ability to capture discriminative receptive fields. In this paper, we design a deformable spatio-temporal attention-based graph convolutional neural network (deformable STA-GCN) that dynamically identifies informative joints and temporal patterns. The proposed framework integrates two key modules, a deformable spatial attention (DeSA) module and a deformable temporal attention (DeTA) module. The DeSA module is integrated to dynamically identify joints closely related to motion, and the DeTA module is incorporated to identify informative temporal receptive fields. Within each module, an attention mechanism is employed to selectively emphasize the sampled features, enabling the network to extract the most discriminative spatiotemporal information. Experimental results on the NTU RGB+D benchmark dataset demonstrate that the proposed model achieves higher accuracy than the comparison model.