Three-dimensional Human Pose Estimation Based On Fully Connected Neural Network

Human pose estimation is an important branch of computer vision,which is widely used in practice,such as intelligent transportation,security monitoring,video entertainment,motion capture and augmented reality.The recognition result of the existing three-dimensional human body pose estimation algorithm is the coordinate position of the three-dimensional joints,and the expression form is mainly the human skeleton composed of the point coordinates or the connection between the points.The expression form of the three-dimensional human body pose is difficult to display more detailed information of the human body,the expression form is abstract,difficult to understand,and does not conform to the expression of the human body pose in the real space environment.In order to solve the above problems,this paper introduces a parameterized three-dimensional human body model(Skinned Multi-Person Linear,SMPL),this model can physically display the three-dimensional pose of the human body.By fitting the three-dimensional human body model with the pose parameter and body shape parameter,the details of the human body can be accurately depicted,which is more realistic than the human body pose represented by joints skeleton.Based on the limited human pose information obtained from single view,this paper proposes to fuse the human pose information from four views.The whole process of three-dimensional human pose estimation is divided into three stages:the first stage is to predict the two-dimensional joints coordinate position of human body from two-dimensional image through hourglass convolution neural network,this network design mainly uses the spatial position relationship between joints,and improves the accuracy of model prediction by fusing the features of different scales;the second stage is the design of three-dimensional human pose regression network,in order to realize the transformation from two-dimensional to three-dimensional space,this paper specially designed a multi-layer cascaded full connection neural network to complete the regression of two-dimensional joints to three-dimensional SMPL model parameters,so that the model can learn the distribution relationship of input and output data,it is necessary to adopt the vector form of joints and three-dimensional model parameters in multi perspective;the last stage is the calculation process of three-dimensional model,according to the model parameters predicted in the previous stage,it is decomposed into the pose parameter and shape parameter of the SMPL model.The three-dimensional SMPL model is generated by the calculation of the correlation function,according to the model,the corresponding three-dimensional joint coordinates can be calculated.The three-dimensional human pose estimation model proposed in this paper takes the two-dimensional joints as the input,forms the three-dimensional human pose regression network model through the cascaded full connection layer,and optimizes the prediction effect by adding batch normalization,dropout and activation function,which improves the efficiency of model prediction and solves the over fitting problem of the network model,which can quickly and accurately estimate the three-dimensional pose of human body.The model proposed in this paper is tested and evaluated in Human3.6M dataset,and compared with the latest similar algorithm,MPJPE is used as the evaluation benchmark.The experimental results show that the average error value of three-dimensional human pose estimation model proposed in this paper under the condition of public protocol is 62.6mm,ranking first in the similar algorithm,compared with Pavlako’s algorithm model,the error is reduced by 9.3mm.Under the same condition,compared with the algorithm of predicting three-dimensional SMPL model,the error value calculated by the model in this paper is lower than the results of most algorithms,ranking in the forefront of many algorithms.In order to verify the efficiency of this model,the running time of this model is evaluated experimentally under the condition of existing equipment,the experimental results show that the forward propagation time of the proposed model is 1.3 seconds,which is more efficient than that of the similar HMR model.Because the network structure of this algorithm is simple,it achieves high accuracy and operation efficiency by directly returning to three-dimensional human body model from two-dimensional joints,and the effect of this algorithm design is confirmed by a number of evaluation benchmark.