分段时间注意力时空图卷积网络的动作识别

Action Recognition with Segment Temporal Attention Spatiotemporal Graph Convolutional Networks

得益于图卷积网络(GCN)对于处理非欧几里得数据有着非常好的效果,同时人体的骨骼点数据相对于RGB视频数据具有更好的环境适应性和动作表达能力.因此,基于骨骼点的人体动作识别方法得到了越来越多的关注和研究.将人体骨骼建模为时空图形的数据进行基于GCN模型的动作识别取得了显著的性能提升,但是现有的基于GCN的动作识别模型往往无法捕获动作视频流中的细节特征.针对此问题,本文提出了一种基于分段时间注意力时空图卷积骨骼点动作识别方法.通过将数据的时间帧进行分段处理,提取注意力,来提高模型对细节特征的提取能力.同时引入协调注意力模块,将位置信息嵌入注意力图中,这种方法增强了模型的泛化能力.在NTU-RGBD数据集和Kinetics-Skeleton数据集上的大量实验表明,本文所提模型可以获得比目前多数文献更高的动作识别精度,有更好的识别效果.

Benefit by the graph convolutional network(GCN)has very good effect on processing non-Euclidean data,and the human skeleton point data has better environmental adaptability and action expression ability than RGB video data.Therefore,human action recognition methods based on skeleton points have received more and more attention and research.Human skeletons are modeled as spatiotemporal graphics data for action recognition based on GCN models,which has achieved significant performance improvements.However,existing GCN-based action recognition models generally unable to capture detailed features in action video streams.Aiming at this problem,this paper proposes a convolutional skeletal point action recognition method based on segmental temporal attention spatiotemporal graph.According to segmenting the time frames of the data and extracting attention to improve the ability of extracting detailed features.At the same time,a coordinated attention module is introduced to embed the location information into the attention map,which enhances the generalization ability of the model.Extensive experiments on NTU-RGBD dataset and Kinetics-Skeleton dataset show that the model proposed in this paper can achieve higher action recognition accuracy and better recognition effect than most current literatures.