动作(物理)
计算机科学
医学
鉴定(生物学)
动作识别
人工智能
背景(考古学)
特征(语言学)
神经科学
疾病
作者
Xing Li,Yangyang Li,Jiajie Lin,Ge Gao,Peng Fei Huang,Hengyi Ren,Yiming Wang,Xin Li,Qian Huang
出处
期刊:
[Association for Computing Machinery]
日期:2026-11-05
卷期号:: 5427-5436
标识
DOI:10.1145/3767308.3836345
摘要
The application of Graph Convolutional Networks (GCNs) has significantly advanced skeleton-based action recognition in the field of computer vision in recent years. However, most existing studies overlook the issue of spatio-temporal interference under non-ideal condition in real-world scenarios, which is widespread and primarily manifested as spatial joint anomalies and temporal frame missingness, leading to substantial degradation in recognition performance. To address these challenges, we propose a Spatio-Temporal Complementary Repair Graph Convolutional Network (STCR-GCN), designed to enhance the robustness of skeleton-based action recognition. Specifically, a Static-Dynamic Synergistic Spatial Repair (SDSR) module is proposed, which integrates static posture structure anchoring and dynamic motion rheology perception to respectively preserve spatial structural integrity and motion continuity, thereby achieving spatial-level repair by synergistically compensating for missing joints and suppressing abnormal jitter. Furthermore, a Rhythm-Driven Temporal Repair (RDTR) module is introduced, which constructs rhythmic semantics through rhythmic primitive decomposition and activates rhythm expert compensation pathways for adaptive restoration of diverse action rhythms. Each expert incorporates dual-attention compensation to effectively address rhythm fragmentation and misalignment, thereby accomplishing temporal-level repair. Experimental results demonstrate that STCR-GCN achieves state-of-the-art performance on various non-ideal datasets, while also delivering excellent results on three mainstream benchmark datasets under ideal conditions. All source code are made available at https://github.com/XingLi1012/STCR-GCN.
科研通智能强力驱动
Strongly Powered by AbleSci AI