| 摘要: |
| 针对人体动作识别领域中因模型结构复杂化引发的计算开销大、运行性能受限的问题，提出一种基于自适应图卷积的语义引导动作识别方法。该方法通过自适应图卷积优化拓扑结构，构建长距离关节间的联系，从而增强模型对骨架数据的特征提取能力。同时，引入轻量化多模块注意力机制，从时间、空间和通道三个维度提高对关键帧和关节的关注度，提高动作识别准确率。此外，通过引入“关节类型”和“帧索引”语义信息，能够有效描述骨架数据中的关节间关系和帧间关系，显著降低了网络参数量。实验结果表明，在NTU RGB+D 60数据集中按照CS划分方法下，识别准确率在现有网络模型中处于较高水平，同时其参数量远低于同等性能的网络模型，有效平衡了模型精度与计算效率。 |
| 关键词: 动作识别 自适应图卷积 语义引导 注意力机制 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Semantic-guided action recognition method based on adaptive graph convolution |
|
Lu Meichen1, Jin Ming2
|
|
1.Shanghai Normal University;2.Shanghai Normal University Tianhua College
|
| Abstract: |
| To address the challenges of high computational cost and limited runtime performance caused by the increasing complexity of model structures in human action recognition, this paper proposed a semantic-guided action recognition method based on adaptive graph convolution. The method optimized the topological structure through adaptive graph convolution, establishing long-range dependencies between joints to enhance the model's ability to extract features from skeleton data. Additionally, a lightweight multi-module attention mechanism was introduced to improve focus on key frames and joints across temporal, spatial, and channel dimensions, thereby enhancing recognition accuracy. Furthermore, by incorporating semantic information such as "joint type" and "frame index," the method effectively captured inter-joint and inter-frame relationships in skeleton data, significantly reducing the number of network parameters. Experimental results on the NTU RGB+D 60 dataset under the CS protocol demonstrated that the proposed method achieved competitive accuracy among state-of-the-art models while maintaining a substantially lower parameter count compared to models with similar performance, effectively balancing model accuracy and computational efficiency. |
| Key words: action recognition adaptive graph convolution semantic-guided attention mechanisms |
