Research On Human Pose Estimation Algorithm Based On Deep Learning

Human pose estimation is a very basic research direction in the field of computer vision.This paper firstly analyzes the research status and existing research difficulties of human pose estimation,and then selects 2D multi-person human pose estimation as the research focus of this paper.The main research contents are as follows:1.The research focus of this paper is to use deep learning technology to solve the problem of 2D multi-person human pose estimation.In order to facilitate the follow-up research,the basic knowledge of deep learning human pose estimation and neural network is firstly introduced.2.In order to efficiently solve the problem of 2D multi-person human pose estimation,a model with high accuracy and fast prediction speed was obtained.Firstly,the top-down algorithm of 2D multi-person human pose estimation algorithm is studied,the network model of high resolution human pose estimation is selected,its basic principle is studied,and the attention mechanism and knowledge distillation are used to improve it.The attention mechanism is introduced into the original high-resolution network model,and the concrete implementation is completed by non_local module,which can enhance the correlation between pixels and other pixels,so as to improve the model's ability to distinguish local information and improve the model's accuracy.In addition,the knowledge distillation method is used,which includes two parts: teacher network and student network.The backbone models of teacher network and student network are both high-resolution human pose estimation networks improved by attention mechanism.Through knowledge distillation,the student network can obtain useful information from the teacher network,so as to obtain a model with fewer parameters and higher accuracy,which improves the feasibility of model implementation.However,when testing the speed of the model,it is found that the real-time performance of the model can be further improved.In addition,as the number of people in the picture increases,the prediction time of the model will also increase significantly.3.In order to efficiently solve the problem of 2D multi-person human pose estimation and reduce the impact of the number of people in the picture on the model's prediction time,the bottom-up algorithm of 2D multi-person human pose estimation algorithm was studied.The model of composite field human pose estimation network is selected and its basic principle is studied.In addition,the network is improved by using dilation convolution and shuffleNet V2 network.First,the shuffleNet V2 network is used as the backbone network to replace the ResNet backbone network in the original composite field pose estimation model,which can improve the prediction speed of the model.In addition,the dilation convolution is introduced into the composite field human pose estimation network model using shufflenet V2 network as the backbone network,so as to increase the receptive field of the network and improve the accuracy of the network.Experiments show that the model has a good accuracy and speed,and the model's prediction time is less sensitive to the number of people in the picture.In addition,for low resolution images,the algorithm still has a certain effect.