Scholarlink

<p>Aiming at the problem that existing graph convolution-based human skeleton action recognition methods struggle to adaptively capture the dynamically changing semantic relationships between human body joints, leading to limited representation capability for non-rigid human motions, a Semantic-Aware Deformable Graph Convolution Network (SAD-GCN) based anomaly behavior recognition method is proposed. Firstly, utilizing semantic information between joints, the similarity of joint features is calculated to select key joints, dynamically adjusting the sampling positions of the graph convolution operation to construct the semantic-aware deformable graph convolution, adaptively eliminating redundant connections and improving action recognition accuracy. Secondly, an Efficient Channel Attention (ECA) module is introduced to differentiate weight disparities between channels, highlighting important channel features. Finally, to fully leverage temporal and spatial information, a Spatio-temporal Adaptive Channel Aggregation (ST-ACA) strategy is employed to generate temporal dynamic weights, which are then multiplied element-wise with spatial channel features, achieving the calibration of spatial weights by temporal dynamics, thereby effectively integrating spatio-temporal information for complex actions. The proposed method achieved accuracies of 94.1% (X-Sub) and 88.7% (X-Set) on the public datasets NTU RGB + D 60 2D and NTU RGB + D 120 2D, respectively, and accuracies of 96.7% (X-Sub) and 99.0% (X-View) on the self-built anomaly behavior dataset. The proposed method outperforms other mainstream methods, validating its effectiveness.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

SAD-GCN: Semantic-aware deformable graph convolutional networks for human abnormal behavior recognition

  • 期刊论文

  • 发表日期：

    2025年11月20日

查看全文
  • Jiakang Dai,
  • Yuping Feng,
  • Yongping Zhu,
  • Mingliang Huo

摘要

Aiming at the problem that existing graph convolution-based human skeleton action recognition methods struggle to adaptively capture the dynamically changing semantic relationships between human body joints, leading to limited representation capability for non-rigid human motions, a Semantic-Aware Deformable Graph Convolution Network (SAD-GCN) based anomaly behavior recognition method is proposed. Firstly, utilizing semantic information between joints, the similarity of joint features is calculated to select key joints, dynamically adjusting the sampling positions of the graph convolution operation to construct the semantic-aware deformable graph convolution, adaptively eliminating redundant connections and improving action recognition accuracy. Secondly, an Efficient Channel Attention (ECA) module is introduced to differentiate weight disparities between channels, highlighting important channel features. Finally, to fully leverage temporal and spatial information, a Spatio-temporal Adaptive Channel Aggregation (ST-ACA) strategy is employed to generate temporal dynamic weights, which are then multiplied element-wise with spatial channel features, achieving the calibration of spatial weights by temporal dynamics, thereby effectively integrating spatio-temporal information for complex actions. The proposed method achieved accuracies of 94.1% (X-Sub) and 88.7% (X-Set) on the public datasets NTU RGB + D 60 2D and NTU RGB + D 120 2D, respectively, and accuracies of 96.7% (X-Sub) and 99.0% (X-View) on the self-built anomaly behavior dataset. The proposed method outperforms other mainstream methods, validating its effectiveness.