Graph Convolutional Network For Skeleton-Based Action Recognition

As a hot and rising research field,human action recognition has developed rapidly in academia and industry in recent years.Especially in human-computer interaction,video surveillance,automatic driving,virtual reality,etc.,it is widely used.Since human action recognition methods based on video sequences are limited by factors such as computational cost,illumination,and scene changes,skeleton-based action recognition methods have been developed.In the task of skeleton-based action recognition,graph convolutional networks have received a lot of attention and achieved remarkable results.However,the current methods still have problems such as incomplete representation of human skeleton,insufficient spatial modeling ability,and single temporal modeling scale.To this end,this paper proposes two human skeleton behavior recognition methods based on graph convolutional neural networks,explores and optimizes human skeleton representation,and enhances the ability of spatiotemporal modeling.The main contributions of this paper are as follows:(1)A multi-stream dynamic topology graph convolutional network model is proposed.Most of the current methods are calculated based on a fixed topology,and it is difficult to effectively capture the complex nature of the human skeleton graph.In order to be able to learn the dynamic embedding features of the human skeleton,this paper proposes dynamic topological graph convolution and dynamic topological time convolution,which can dynamically generate topological structures according to different behavioral characteristics,so that the key features of actions can be dynamically learned.In addition,combined with the method of multi-information stream input,this paper inputs two different spatial information and two different temporal information into the multi-stream dynamic topological graph convolutional network model for training,and obtains prediction scores and weights respectively.Blend for the final result.Finally,the effectiveness and rationality of the algorithm are verified through a large number of experiments on two large public datasets,NTU-RGB+D and NTU-RGB+D 120.(2)A multi-part adaptive graph convolutional network model is proposed.The human skeleton is composed of different body parts,and actions can be understood as the joint participation of different body parts of the human body.Inspired by this,this work no longer performs graph convolution on the entire skeleton graph,but proposes a multi-part adaptive graph convolution to adaptively learn the topology of each part of the body and dynamically aggregate the correlation between them.Aiming at the problem of single temporal modeling scale,this paper uses an improved multi-scale temporal convolution to improve the temporal dimension feature extraction ability of the model,so as to better learn the temporal and spatial independence and correlation between different parts.Finally,the model is also trained and tested on the above two large-scale public datasets.The ablation experiment verifies the effectiveness of the algorithm.The comparison experiment shows that the multi-part adaptive graph convolutional network improves the accuracy of skeleton-based human action recognition.