Research On Action Recognition Algorithm Based On Multi-scale And Multi Branch Features

With the development of society,mankind has entered the information age.Video has become the main carrier of information and has been used in all fields of life and work.As an application field of artificial intelligence,human action recognition based on video carrier has become the core technology of many display needs,and plays an important role in the fields of intelligent monitoring,human-computer interaction and so on.With the application of depth camera in research and production in human society,the application dimension of action recognition gradually changes from twodimensional to three-dimensional space,that is,action recognition based on 3D human skeleton.Compared with RGB video sequence,3D human skeleton sequence data has the advantages of small amount of data and no interference from the external environment.It soon became one of the key research directions.At present,many 3D human skeleton action recognition methods based on neural network have been proposed.Among them,graph convolution neural network is a research hotspot of action recognition because it can introduce human topology.On this basis,it has also made some development.However,how to provide multi-level and various feature information to the network model through the original data,and how to extract more time features from the method based on graph convolution neural network still need to be further studied.Therefore,this paper studies the above problems.The main work and innovations are as follows:(1)According to the input original human skeleton data,a feature enhancement method using multiple methods is proposed.The improved multi branch fusion feature is input into the network structure,and the multi-level and multi variety feature information is fused by early fusion.(2)Based on multi-scale features,the multi-scale temporal convolution used in this paper is proposed,which is combined with graph convolution to form a multi-scale spatio-temporal convolution module for extracting spatio-temporal features.Finally,according to the above work,the multi branch and multi-scale convolutional neural network structure used in this paper is proposed,sufficient comparative experiments are carried out for the internal parameters of the network,and ablation experiments are carried out with other algorithms in the data set of the mainstream field of human skeleton action recognition to verify the effectiveness of the network used in this paper.