Research On Human Action Recognition Based On Motion Sequence Features

In the past decade,computer vision technology has developed rapidly.Human motion recognition,as one of the key research directions in the field of computer vision,has become a research hotspot in recent years.Human movement recognition is based on the given human skeleton joints and other motion sequences with behavior analysis,using a series of technical methods to analyze and distinguish the specific movement categories.Effectively analyzing and mining the information contained in the characteristics of human motion sequences and quickly and accurately identifying the corresponding motion categories are the key contents of human motion recognition.Because of the strong function of human motion recognition technology,it has a great application prospect in video surveillance and security,medical monitoring,learning and entertainment,autonomous driving vehicles and other directions.Skeleton motion sequence data can provide complete and effective body posture information,and it is insensitive to changes of external conditions such as illumination,and the data volume is small.This paper mainly studies human action recognition algorithm based on motion sequence features.Through in-depth analysis of various model algorithms for human action recognition,this paper proposes a multi-flow graph convolution residual network algorithm based on skeleton sequence data based on spatio-temporal graph convolutional neural network.The research contents of this paper are as follows:1.Aiming at the problem of insufficient action feature extraction from the current mainstream network model,this paper proposes an action recognition algorithm based on the multi-stream network structure,which integrates the information of multiple data streams based on the multi-stream network.Based on the skeleton joint information,the skeleton information is further processed to extract the joint motion information,bone information and bone motion information.For the sake of calculation cost,before data is input into the network framework,this paper integrates the above four kinds of skeleton information to improve the recognition accuracy and effectively reduce the calculation amount.2.To solve the problem of difficulty in modeling skeleton information,an end-to-end graph convolutional network structure based on graph convolutional network is implemented in this paper.Combined with the advantages of spatio-temporal feature extraction of graph convolutional neural network,an adaptive graph convolutional network algorithm is implemented to mine the deep features in the non-directly connected joints of skeleton information,and the spatial attention mechanism is added to give more weight to the joints with rich information.3.Time series information in action recognition is particularly important.In order to extract deeper information in action time series,this paper proposes a cross-domain connection structure,which can mine spatial and temporal features more effectively by introducing cross-domain time graph convolution module unit.4.Inspired by the RESNET network in the convolutional neural network,residual structural units were added into the graph convolutional neural network GCN to reduce the difficulty of model training,increase the stability of the structure while maintaining the accuracy of the model,and improve the robustness of the algorithm.The proposed algorithm model was verified on the NTU-RGB+D data set,and the effectiveness of the proposed method was verified by experimental analysis and comparison with the advanced network model.