Research On Human Behavior Recognition Algorithm Based On V-DBN And Spital Double-stream LSTM

In recent years,deep learning algorithms have made great breakthroughs in many fields,greatly promoting the development of artificial intelligence.Great breakthroughs have been made in the classification,segmentation and target detection of RGB images by various deep learning network models based on convolutional neural network,which inspired researchers to further study the deep learning model.Thanks to the commercialization of depth sensors,it is becoming more and more convenient for people to obtain 3D data.Application programs can extract 3d position information of human body area and key points based on depth information of human body in the scene.By extracting the relevant features of key points,the network model can complete the recognition of human behavior.This subject mainly studies human behavior recognition based on 3D skeleton points,and designs two kinds of network models based on V-DBN and spatial two-stream LSTM.The main research contents and methods are as follows:(1)This paper proposed a method to design and extract spatial-temporal from raw 3D skeleton point location information.In the time dimension,the displacement,velocity and acceleration characteristics of the same skeleton points in the adjacent frames were used to describe changing trajectories.In spatial dimension,the relative position between specified skeleton points and the reference skeleton points in the same frame are used to describe the motion state.(2)This paper applied VLAD(vector of locally aggregated descriptors)method to normalize out of synchronism feature data.Due to the individual differences among different targets,the sequence length between actions is different,which is not conducive to the recognition of actions by classifiers.In this study,VLAD algorithm was used to unify the length of action sequences.(3)This topic designs V-DBN model to complete action recognition.In this model,VLAD algorithm is combined with DBN,feature dimension reduction is completed by contrastive divergence algorithm,and key parameters in the network are analyzed.Besides,this paper analyzes the influence of parameters,combination features,initialization methods,feature fusion methods and different classifiers on recognition results are discussed by experiments.(4)The VLAD algorithm used in the V-DBN model would break the relationship between one frame and neighbor frame.To address this problem,this paper designs spatial two-stream network model to complete the motion recognition.Based on the Tensorflow framework,this model builds a network model through several LSTM network units,aiming at overcoming the difficult of information loss result of VLAD algorithm.In the experiment,a variety of network model construction methods based on LSTM model are compared.This subject mainly completed the spatial-temporal motion feature design based on 3D skeleton points,improved the feature coding and normalization methods,and built two kinds of network structures for human recognition.Experiments based on two common databases show that the proposed method is effective in feature design,feature coding and action recognition.