山西大学学报（自然科学版）2026，Vol.49Issue(2)：220-231,12.DOI:10.13451/j.sxu.ns.2025094
基于自适应通道特征交互融合图卷积网络的骨骼行为识别
Adaptive Channel Feature Interactive Fusion Network for Skeleton-based Action Recognition
摘要
Abstract
This study proposed a novel skeleton-based action recognition method utilizing Graph Convolutional Networks(GCN),which addressed the limitations of conventional spatiotemporal graph convolution frameworks that uniformly process spatiotempo-ral features while neglecting inter-channel interactions.Specifically,the proposed model enhanced the complementary representation of spatial information through the fusion of multiple topological matrices coupled with the introduction of a Channel Interaction At-tention(CIA)module.The CIA module was designed to capture dynamic frame-level information and human structural features across spatiotemporal dimensions,effectively modeling inter-channel relationships and thereby improving skeletal data representa-tion.Furthermore,a Temporal Adaptive Feature Fusion(TAF)module was incorporated to adaptively select varying dilation rates and kernel sizes across network layers.This module replaced traditional residual connections between initial features and temporal module outputs,effectively addressing context aggregation and initial feature integration challenges.The TAF module separately pro-cessed initial features and temporal information,enabling efficient feature fusion and successful integration of initial features with high-dimensional temporal features,which significantly enhanced spatiotemporal feature extraction.Experimental results demon-strated that on the NW-UCLA dataset,the proposed method achieved 2.1％higher recognition accuracy than the baseline model CTR-GCN(Channel-wise Topology Refinement Graph Convolution Network)and 0.7％improvement over state-of-the-art methods Info-GCN.For the NTU RGB+D 120 and NTU RGB+D datasets under different splits,the model showed consistent performance gains of 0.7％,0.8％and 0.5％,0.6％,respectively,surpassing all existing methods across evaluation metrics.These results con-firmed the model's superior performance in both spatiotemporal feature extraction and skeleton-based action recognition tasks.关键词
骨骼行为识别/通道注意力机制/时空特征融合/通道交互Key words
skeleton-based action recognition/channel attention mechanism/spatiotemporal feature fusion/channel interaction分类
信息技术与安全科学引用本文复制引用
施宇航,陈琳琳,郭峰,何强..基于自适应通道特征交互融合图卷积网络的骨骼行为识别[J].山西大学学报（自然科学版）,2026,49(2):220-231,12.基金项目
国家自然科学基金(12301581) （12301581）
北京市自然科学基金(4252033) （4252033）
北京市教育委员会科学研究计划项目(KM202210016002) （KM202210016002）
北京建筑大学基本科研业务费资助(X25039) （X25039）
北京建筑大学硕士研究生创新项目(PG2025172) （PG2025172）
