Research And Application Of 3D Posture Recognition Technology Based On Two Stages

Three-dimensional human posture is an important information point to describe human input instructions.Machine can well read the meaning of human posture,which has a great significance for human-computer interaction.At present,there are some problems in three-dimensional human pose recognition,such as generated action sequence jitter,slow recognition speed and inaccurate detection during environmental occlusion.This paper studies the above problems and proposes the improvement strategy,and finally makes a real-time 3 D human pose recognition system according to the improved content.This paper focuses on three-dimensional human attitude recognition technology,combined with the existing technology,analyze the shortcomings,puts forward the TCN network and Open Pose two-dimensional attitude detector,and use angular vector calculation reconstruction three-dimensional human coordinates,finally through the model compression technology to improve the detection speed,finally in the experimental verification and build three-dimensional attitude recognition system.The experimental results and systematic demonstration show that the 3 D human pose recognition method has practical application value in virtual creation and art performance.The main research contents are as follows:We propose a 3D human pose recognition method combining Open Pose,TCN network and angular vector calculation.Initial 2D human pose recognition was completed using Open Pose,and 2D human bodies were further processed using the TCN network to eliminate temporal jitter and generate smooth action sequences.Finally,the twodimensional human pose is reconstructed into three-dimensional dimensions using angular vector calculations.Experiments show that this two-stage 3D human pose recognition method can correctly identify 3D human coordinate points in occlusion,blur and other scenarios,and perform well in Human Eva and Human 3.6M datasets.A compression method of human posture model based on channel redundancy is proposed,which can effectively improve the detection speed of human posture recognition model.By analyzing the feasibility of the compressed model parameters,the algorithm is designed to detect the redundant channels in the model and speed up the model inference.Extensive experiments show that the proposed method improves the clarity of the joint point heatmap while squeezing the model volume.This thesis constructs a 3 D human pose recognition system composed of four modules:data acquisition,human body detection,pose recognition and model output,and integrates the proposed algorithm.Then,the construction of the system was completed under the corresponding hardware environment,and the functions of three-dimensional human pose recognition and output model are displayed through experiments.