Research On Human Pose Estimation Algorithm For Outdoor Images

Human pose estimation based on RGB image refers to detecting the position of each part of the human body from the image and calculating its direction and scale information.which can be divided into two cases: 2D and 3D.The 2D human pose estimation aims to get the coordinates of joint key points in the 2D image;while the 3D human pose estimation needs to calculate the depth information of the joint point except for estimating its 2D position coordinates,and finally obtain the posture information of the human body in 3D space.However,due to the difficulty of marking data,in the field of human pose estimation,the datasets in outdoor scenes only have 2D coordinate information marked,while the datasets with 3D information marked are all collected under indoor restricted scenes.Therefore,in order to solve the above problems,this paper refers to the 3D human pose estimation method based on the weakly supervised migration learning for outdoor images.Taking the dataset with 2D outdoor complex information and the dataset with 3D indoor information as training data,we train and get an end-to-end 3D human pose estimation model to reconstruct the 3D pose model of the human body in complex outdoor scenes.We study the basic principle and process of the weakly supervised migration learning method and train the model for several times.This model has poor robustness for different scenarios and actions.Not only that,if we repeatedly train the model with the same training parameters,the experimental results show that the convergence curve of the model has large fluctuation.What's worse,each convergence curves are different and each performance of training the model is also quite different.To solve the above problems,by visualizing the output characteristics of each convolutional layer,it has found that whether the extracted features are “universal” has a great influence on the performance and robustness of the migration learning method.So,to improve the accuracy and robustness of model,this paper proposes a Self-Paced Transfer Learning method based on Self-Paced Learning.Which learns the sample space from "easy to difficult,fast and then slow",and overcomes the problem of poor robustness of the model and improve its accuracy to a certain degree.For the stacked hourglass model,which aims to solve the problem of 2D pose estimation and the weak depth information estimation performance.This paper introduces the structure of the SENet module and the feature weight recalibration module.These modules make the front and back features interact with each other,which not only enhances the feature space expression ability and provides the feature maps with more depth information to the subsequent depth estimation sub-models,but also enhance the correlation of the front and back sample features to improve the stability of the model.In addition,the bilinear interpolation method is adopted in the up-sampling of the network,which enhances the performance of the network features and provides a good foundation for the depth estimation.This paper mainly uses MPII and Human3.6M public datasets for training and uses Human3.6M and MPI-INF-3DHP public datasets for test.The experimental results verify that get the effectively improve by our method to the baseline.The MPJPE on the Human 3.6M dataset is 60.69 mm,and PCKh is 92.3%,4.21 mm and 0.7% are improved relative to the benchmark model,respectively.The MPJPE on the MPI-INF-3DHP dataset is 41.35 mm,and the PCKh is 90.84%,17.35 mm and 6.22% are improved relative to the benchmark model,respectively.