基于骨架的渐进式模式重构与阶段式协同对比人体行为识别研究

A Study on Skeleton-Based Human Action Recognition via Progressive Pattern Reconstruction and Stage-Wise Collaborative Contrastive Learning

摘要: 骨架动作识别旨在从人体关节序列中学习具有判别性的时空表示。针对现有方法拓扑先验易弱化、相似动作细粒度差异建模不足以及时间维区间协同约束缺乏等问题，本文提出一种面向骨架动作识别的渐进式模式重构与阶段式协同对比方法。针对拓扑先验弱化，引入样本自适应骨骼先验融合策略，将样本级几何先验注入拓扑学习，以缓解原始骨骼语义偏离；针对细粒度差异建模不足，设计渐进式模式重构模块，两阶段串联捕获主导模式并补充局部判别信息；针对区间协同约束缺乏，构建阶段式协同对比损失，在时间轴划分阶段区间并施加监督对比，以增强类内紧凑性与类间可分性。实验结果表明，所提方法在NTU RGB + D60、NTU RGB + D120和NW-UCLA等主流数据集上取得了具有竞争力的识别性能，消融实验进一步验证了各组成模块的有效性。结果表明，所提方法能够有效增强骨架拓扑建模和细粒度动作表示能力。

Abstract: Skeleton-based action recognition aims to learn discriminative spatiotemporal representations from human joint sequences. Existing methods still face three challenges: weakened topological priors, insufficient modeling of fine-grained differences between similar actions, and limited interval-level collaborative constraints along the temporal dimension. To address these issues, this paper proposes a progressive pattern reconstruction and stage-wise collaborative contrastive learning method for skeleton-based action recognition. First, a sample-adaptive skeletal prior fusion strategy injects sample-level geometric priors into topology learning to alleviate deviations from the original skeletal semantics. Second, a progressive pattern reconstruction module employs a two-stage reconstruction process to capture dominant patterns and complementary local discriminative information. Third, a stage-wise collaborative contrastive loss divides the temporal axis into multiple stage intervals and applies supervised contrastive constraints to improve intra-class compactness and inter-class separability. Experimental results on NTU RGB + D60, NTU RGB + D120, and NW-UCLA demonstrate that the proposed method achieves competitive recognition performance, and ablation studies further verify the effectiveness of each component. The results show that the proposed method can effectively enhance skeletal topology modeling and fine-grained action representation.