| Skeleton-based action recognition aims to study how computers extract motion pattern features for action classification through the time sequence input of human skeleton nodes.It has a wide range of applications in human-computer interaction,intelligent monitoring systems,unmanned security,patient monitoring,and motion sensing games.Graph convolution,due to its powerful representation ability for graph data,has been widely used in skeleton-based action recognition tasks,and many excellent research results have emerged in this field over the years.However,existing graph convolution methods still have some drawbacks,such as using shared graph topology for all frames and channels during feature aggregation,modeling all categories of actions with a fixed-scale topology,modeling keyframes and non-keyframes uniformly,and having too many model parameters.These drawbacks greatly limit the representation ability and engineering applications of graph convolution networks.To address the issues mentioned above,this paper proposes methods such as specific topology learning,multi-scale topology modeling,and multi-scale keyframe enhancement to improve the representation ability of Graph Convolutional Networks(GCNs)and integrate multi-scale spatio-temporal features,thereby enhancing the network’s classification ability for actions with diverse motion patterns and making the network more lightweight.The main research contents of this paper are outlined as follows:(1)The background,significance,existing technical challenges,and technical framework of skeleton-based action recognition task are first elaborated.Then,the current research status of skeleton-based action recognition task is summarized and reviewed from the perspective of deep learning architecture.Finally,a novel graph convolution representation evaluation theory is used to analyze and compare Graph Convolutional Network(GCN)-based skeleton action recognition methods.(2)To address the limited representation ability of graph convolutional networks(GCN)caused by the shared graph topology used for aggregating features across all frames and channels,this paper proposes Multi-dimensional Dynamic Topology Learning Graph Convolution Network(MD~2TL-GCN)to dynamically learn the specific topology for each input dimension,and cleverly carry out the interactive learning of multi-dimensional topology.MD~2TL-GCN can model the topology with temporal and channel specificity at the same time,so as to obtain comprehensive and effective representative features.MD~2TL-GCN achieves 92.64%and 89.29%accuracy on NTU-RGB+D and NTU-RGB+D 120 datasets,respectively,which is superior to the current state-of-the-art methods.(3)To address the problems of existing GCNs in spatio-temporal modeling such as inability to distinguish keyframes,inability to perform multi-scale topology learning,and lack of lightweight models,this paper proposes a lightweight Multi-scale Spatio-Temporal Dynamic Graph Convolution Network(MSTD-GCN).During topology modeling,it can construct multiple scales of topology to match different motion patterns by screening nodes based on their degree of participation.During temporal modeling,the network enhances its temporal modeling capability by highlighting keyframes.Similarly,experimental analysis was performed on two public datasets,and the results showed that MSTD-GCN achieved comparable accuracy to current state-of-the-art methods on two large-scale datasets,NTU-RGB+D and NTU-RGB+D 120,while the model parameter size was one order of magnitude smaller than current methods. |
