Multi-person Pose Estimation Based On Convolutional Neural Network

Human pose estimation is an important research direction in the field of computer vision.It is used to detect the position of several key points of the human body from images or videos to help the machine to better interpret human behavior.Human pose estimation is the basis of subsequent tasks such as human-computer interaction,human body movement recognition,and pedestrian re-recognition.It has important application value in the fields of automatic driving,somatosensory games,video surveillance,sports training,and auxiliary medicine.With the continuous development of deep learning,deep neural networks have made breakthrough progress in many sub-fields such as image classification,semantic segmentation,and target detection in computer vision.More and more researchers have applied deep neural networks to human pose estimation.In this paper,aiming at the problems in the multi-person pose estimation method in static pictures and videos,we studied and improved the existing convolutional neural network algorithm model,and proposed a multi-person pose estimation method based on bi-directional weighted fusion and a fast multi-person pose estimation method based on depthwise separable convolution.The main work includes:(1)In terms of multi-person pose estimation in static pictures,a multi-person pose estimation method based on bi-directional weighted fusion is proposed.Aiming at the problems of insufficient utilization of feature maps and low estimation performance of existing Cascaded Pyramid Network(CPN)models,this paper proposes a multi-person pose estimation method based on bi-directional weighted fusion.The main improvements of this method over the CPN model include: first,using a bi-directional weighted feature pyramid network to improve Global Net so that feature maps of different scales in the feature pyramid are fused according to different proportions,and in addition to the original top-to-bottom connection method,the feature pyramid is added from bottom to top.This way,the feature maps of different scales are effectively used;the second is to add the network connection method in Dense Net to Refine Net,so that new feature maps obtained after each layer of convolution are concatenated with the original input feature maps instead of adding,so that after multiple convolutional layers,the feature maps can be reused multiple times,which can increase the diversity of features,and enable the network model to learn more information,thereby effectively improving the estimation accuracy.(2)In the aspect of multi-person pose estimation in the video,an improved Open Pose model based on a depthwise separable convolution structure is proposed.Because the traditional Open Pose model relies on a convolutional neural network with high computational complexity,the training and deployment of the model requires a server equipped with a high-performance GPU board to complete it.It cannot be put into practical use in edge computing using embedded devices.In response to this problem,this paper replaces the traditional convolutional structure in the Open Pose model with a depthwise separable convolution structure,and modifies the number of repeated cascade stages in the Open Pose model,thereby proposing an improved Open Pose model based on the depthwise separable convolution structure.Under the condition of keeping the estimation accuracy comparable to the original Open Pose model,the calculation amount of the Open Pose model is greatly reduced,so that the improved model can realize real-time operation on embedded terminals with less computing resources,and supports real-time human pose estimation in edge computing scenarios applications.