ASR-GCN: Adaptive spatial information reconstruction GCN for skeleton-based action recognition

Over the past few years, skeleton-based action recognition has attracted significant focus in the area of computer vision. However, existing methods still face many challenges in feature extraction and dynamic feature learning for complex tasks. This paper proposes an innovative Adaptive Spatial Information Reconstruction Model (ASR-GCN) to address these challenges. By optimizing the feature extraction process, the recognition capability of the model is significantly improved while introducing only a minimal increase in parameters and computational complexity. First, we propose a Gated Reconstruction Unit (GRU), which promotes the model's learning of representative features by employing a reweighting gating strategy, significantly improving recognition efficiency. Furthermore, we observed that existing methods generally relied on fixed feature extraction strategies, which limits the deep mining of intrinsic features of skeleton data. Therefore, we constructed an Adaptive Spatial Information Reconstruction Unit (ASRU) to adaptively adjust the contribution of features, efficiently extract, reweight, and integrate feature information, enhancing the feature extraction capability of the model. We conducted experiments on two large action recognition benchmark datasets, NTU RGB+D 60 and NTU RGB+D 120. The results show that our model achieves current state-of-the-art performance on both benchmarks. On the cross-subject and cross-set of the NTU 120, the accuracy reaches 90.9% and 92.4%, respectively.