Research On Human Pose Estimation Based On Convolutional Neural Network

Human pose estimation is an important research content in computer vision,which is widely used in video surveillance,entertainment industry,motion analysis and other fields.2D human pose estimation is mainly divided into single human pose estimation and multi-person human pose estimation.This paper mainly studies 2D multi-person human pose estimation.Convolutional neural network is an important part of deep learning.At present,most 2D pose estimation methods use convolutional neural network.In recent years,human pose estimation mainly focuses on designing better deep network model to extract more abundant features,so as to improve the accuracy of key point prediction.In the research of the latest human attitude estimation model,the high resolution convolutional neural network HRNet provides a new direction for this field due to its special parallel structure.In this paper,HRNet network is selected as a tool to study 2D human pose estimation.The specific research contents are as follows:(1)In this paper,attention mechanism(REFAM)integrating different resolutions is introduced to improve the fusion method of features.In the process of feature extraction,effective information is activated and useless information is suppressed,so as to improve the overall accuracy of human key point detection.(2)This paper introduces the void space convolution(ASPP)module,which has a good effect in the target detection task.On the basis of not reducing the resolution of the feature map,the feature field is increased to improve the positioning accuracy of the small-sized human key points.Higher semantic information can also be extracted through the ASPP module,which is beneficial to key point prediction.(3)This paper introduces the intermediate supervision mode from rough to fine to improve the supervision mode of HRNet,alleviate the problem of gradient disappearance,and improve the predictive ability of network model.The above three research contents were trained,verified and tested on COCO sets.The experimental results showed that the introduction of attention mechanism integrating different resolutions,void spatial convolution and intermediate supervision mode from thick to thin all improved the final results of the model.