| Estimation three-dimensional human pose from the position of two-dimensional joints has shown results.In the recent,most methods use deep learning models and use two-dimensional joint point coordinates to predict the depth value of the three-dimensional coordinates.The acquisition of two-dimensional joint point coordinates is based on a more accurate two-dimensional coordinate detector.However,using 2D joint coordinates as input loses more information than image-based approaches and results in ambiguity.In order to overcome this problem,we combine bone length and camera parameters with 2D joint coordinates for input.This combination is more discrimination than the 2D joint coordinates in that it can improve the accuracy of the model’s prediction depth and alleviate the ambiguity that comes from projecting 3D coordinates ground truth and the output of the proposed model.In addition,we have introduced direction constraints to make the prediction results of the deep learning model more accurate.The main research works of this paper are as follow:(1)This thesis proposes ExtraPoseNet which can combine the bone length and camera parameters.Under the action of camera parameters and bone length,the model’s discriminative ability to data has been significantly improved,making it easier to predict and more accurate the three-dimensional coordinate value.The problem of information loss based on two-dimensional coordinates as input is alleviated.(2)This algorithm designs bone direction constraints(BDC)based on joint points to make up for the inaccuracy of single mean square error(MSE)loss prediction results.It enables the model to introduce the calculation of the bone orientation in the prediction process to obtain better depth prediction results.(3)The algorithm ExtraPoseNet tested the model on two authoritative datasets Human3.6M and MPI-INF-3DHP.Experimental results show that Extra Pose Net is helpful for learning the mapping relationship between 2D joint points and 3D joint coordinates,and can predict the depth value of joint points more accurately.Compared with the mainstream 3D human pose estimation methods,Extra Pose Net has achieved very competitive results.It has achieved P1(MPJPE)48.8 and P2(PA-MPJPE)39.0 prediction results on the public authoritative dataset Human3.6M test set. |
