3D Human Pose Recognition Via Deep Learning

Human body pose recognition has a wide range of application scenarios,such as human-computer interaction,virtual reality,and medical diagnosis.At present,a lot of research work is based on pictures,videos or skeleton representations of the human body.With the maturity of 3D scanning and modeling technology,more and more 3D human body datasets appear.Static pose recognition,retrieval and dynamic pose recognition in 3D human pose recognition are important research topics in computer graphics.The pose recognition based on the 3D human body model has become an urgent problem to be solved.In response to the above problems,the research content of this thesis mainly includes the following points:(1)A 3D human body static pose recognition method based on multi-view convolutional neural network is proposed.First,set up a set of virtual cameras to project the 3D human body model.Secondly,the Ordered View Feature Fusion(OVFF)algorithm is proposed to replace the View-pooling layer in the multi-view convolutional neural network to fuse the multi-view features,effectively solving the problem of directional correlation recognition of the 3D human body model.The above methods are evaluated by SH-RE,SH-SY,FAUST and HPRD datasets,and the accuracy rates are 100%,99.38%,88.33% and 97%,respectively.(2)A method of 3D body static pose retrieval based on Siamese network is proposed.First,a pair of 3D human body models are respectively mapped to new feature vectors through the Siamese network.Secondly,the Euclidean distance between the output vectors is calculated through the loss function.The more similar the pose between the two mannequins,the smaller the Euclidean distance.The greater the difference in pose between the mannequins,the greater the Euclidean distance.The Siamese network is composed of two identical multi-view convolutional neural networks proposed in static pose recognition.Under the Nearest Neighbour(NN),First Tier(FT)and Second Tier(ST)evaluation indicators,the algorithm in this paper is improved by 9.3%,12.8% and 11.1%,respectively,compared with traditional feature extraction methods.(3)A 3D human body dynamic pose recognition method based on Long Short-Term Memory(LSTM)network is proposed.First,the first frame model in the3 D human action sequence is selected as the sequence template,and the shape difference of the subsequent model of the action sequence relative to the template is calculated by the shape difference operator,and it is expressed as a low-dimensional shape difference information tensor.Then,through the combination of 2D convolutional neural network and LSTM,the spatial and temporal dimensional features of the shape difference information tensor are extracted,so as to realize the recognition of the dynamic pose of the human body.The above methods are evaluated by the dynamic pose dataset Human Eva,Mo Sh,SFU,SSM and Transitions,and the classification accuracy is 98.4%,99.7%,100%,99.4% and 100%,respectively.