基于骨架特征的扩展可分离层多尺度图卷积动作识别方法

EpSepLayer multi-scale graph convolutional network for skeleton-base action recognition

随着动作识别任务在计算机视觉领域的飞速发展,更具信息密度的骨架数据已然成为动作表示的重要形式。针对骨架数据动作识别任务的中时空建模问题,提出一种自适应的扩展可分离层多尺度图卷积网络。在空间模块中引入自适应模块,通过网络迭代得到可训练的邻接矩阵,更好地适应各种动作,增强模型灵活性;在时间模块中,采用包含可分离卷积结构的多尺度时序卷积,在提高时间建模能力的基础上,减少计算复杂度和参数。在NTURGB+D60和NTURGB+D120数据集上的实验验证了算法的有效性和先进性。

With the rapid advancement of action recognition tasks in the field of computer vision,skeleton data,which offers rich-er information density,has emerged as a vital form of action representation.To address the challenge of building spatial-temporal mod-els for skeleton-based action recognition,we propose an adaptive expanded separable multi-scale graph convolutional network.The adaptive module is introduced in the spatial module to obtain a trainable adjacency matrix through network iteration,which can better adapt to various actions and enhance model flexibility.In the temporal module,a multi-scale temporal convolution containing a separa-ble convolution structure is used.On the basis of improving time modeling capabilities,computational complexity and parameters are reduced.Experimental results on large-scale skeleton action recognition datasets,NTU RGB+D 60 and NTU RGB+D 120,demonstrate that our algorithm outperforms state-of-the-art methods.