Research On Human Pose Estimation Based On Multi-dimensional Feature Fusion And Collaborative Feature Enhancement

As a basic task in the field of computer vision,human pose estimation has a great influence on other tasks in the field of computational vision.Related research work has applied deep learning algorithms to human pose estimation tasks,which has improved human pose estimation to a certain extent.accuracy.However,due to the incomplete feature extraction dimension and the insufficient transmission of cascade structure information,the accuracy of the human pose estimation model has been restricted to some extent.In view of the above problems,this dissertation models from the two perspectives of the internal network design of each module in the model and the feature transfer between each module in the model,and proposes a cascade network model based on multi-dimensional feature fusion and a weighted collaborative feature enhancement level.Hourglass model.First,for the problem of incomplete feature extraction dimensions in existing models,the angle of constructing the model is concentrated in the network design of each module,fully considering the characteristics of multi-scale features and multi-receptive field feature information,a multi-dimensional feature fusion based Cascading network model.In this model,the multi-dimensional feature extraction network and the hourglass network are constructed in parallel to form a multi-dimensional feature fusion network,and the multi-dimensional feature fusion network is cascaded to finally realize the prediction of the joint points of the actors in the image.Secondly,in view of the problem of inadequate information transmission of cascade structure,the angle of constructing the model is focused on the feature transmission between network modules,fully considering the consistency of feature information at the same scale,a weighted collaborative feature enhancement cascade hourglass is proposed The model builds a collaborative feature enhancement method from the perspective of feature transfer between network modules,and weights the output results of different stages as the final output of the model,which improves the accuracy of the prediction results of the joint points of the actors in the image.degree.Finally,the two human pose estimation models designed from different angles are experimentally verified on the MPII data set and the LSP data set.Experiments prove that the models constructed in this dissertation can effectively improve the accuracy of the model for the prediction of joint points.