Research And Implementation Of 3d Human Pose Recognition Algorithm Based On Residual Network

Human pose recognition is fundamental and challenging in the field of computer vision.Compared with the two dimensional(2D)human pose recognition,the three dimensional(3D)human pose recognition can be more accurately applied to scenes such as abnormal behavior detection,3D animation rendering and robot action teaching in the three-dimensional space because of the depth information it contains.Current 3D human pose recognition methods are mostly based on residual networks,which implement feature extraction through multi-layer convolution on RGB images,and then use deep feature maps to obtain human joint heat maps,and finally extract joint coordinates.However,these methods often only rely on deep feature maps,and do not enhance the feature information,the error of pose recognition is large.Therefore,this paper explores the enhancement of feature information with the help of feature fusion and attention mechanism based on the residual network.The main research contents are as follows:(1)A method based on equalization and fusion of multi-scale feature maps is proposed.This method firstly equalizes the number of channels and sizes of the deep and shallow feature maps,and secondly fuses feature maps of different scales to complete the enhancement of feature information.On the Human3.6M dataset,compared with previous methods,the mean root position error is reduced by 0.9 mm,and the PA mean per joint position error is reduced by 4.5 mm.Experiments show that the equalization fusion method can effectively reduce the joint error of 3D human pose recognition.(2)A method based on the cascade fusion of multi-scale feature maps is p roposed,which fuses multiple feature maps from deep to shallow.Compared w ith the equalization fusion method,this method uses the progressive relationship between different depth feature maps to carry out cascading fusion layer by l ayer,which has the characteristics of clear and orderly structure and progressiv e fusion.On the Human3.6M dataset,compared with previous methods,the me an root position error is reduced by 1.2 mm,and the PA mean per joint positi on error is reduced by 5.2 mm.The experimental results show that the cascad e fusion method works excellent.(3)A two-branch network method based on attention mechanism is proposed.Compared with multi-scale feature map fusion method,the method branch one uses the residual network to learn the feature information of the pose,the second branch uses the attention mechanism to learn the attention distribution information of the pose,and then the method uses the attention distribution to enhance the feature information.On the Human3.6M dataset,compared with previous methods,the mean root position error of the model is reduced by 1.7 mm,and the PA mean per joint position error is reduced by 8.2 mm.The experimental results show that the method can obtain more accurate human pose.