Research On Human Pose Estimation Algorithm And Its Lightweight Based On Deep Learning

Human pose estimation is one of the research focuses in the field of computer vision,and it has become more and more important in practical applications,such as human-computer interaction,video surveillance,and animation production.However,convolutional neural networks with large parameters and large calculations are not conducive to implementation on mobile terminals or embedded devices with limited storage.In order to make the human pose estimation network model have better practical applications,it is necessary to design lightweight model.At the same time,in order to strengthen the learning ability of the human pose estimation network and improve the detection performance,it is necessary to conduct in-depth research on the extraction and fusion of multi-scale and multi-level features in the network.Based on the pyramid hourglass network in human pose estimation,this paper designs a lightweight network and a multi-level high-resolution feature fusion network.The details are as follows:This paper proposes a lightweight pyramid hourglass network for the classic pyramid hourglass network.It reduces training parameters by using depth-wise separable convolution instead of ordinary convolution.At the same time,channel split module and shuffle module are added to change the channel dimension of feature map to reduce the amount of parameters and strengthen the fusion of features.Tested on the standard MPII dataset,experiments show that the lightweight pyramid hourglass network effectively reduces the network parameters and the computational complexity.It reduces the parameter storage space by approximately 50% and maintains considerable accuracy.Aiming at the shortcomings of high-resolution feature extraction and fusion in hourglass networks,a multi-level high-resolution feature fusion network is proposed.This article focuses on the extraction of high-resolution features because they have the richest information.In the process of extracting features of different resolutions,the network continuously supplements high-resolution features.In addition,a depth pyramid residual module is designed to fuse features of different levels.The entire network is generated by stacked sub-networks.In order to apply in the limited storage space better,this article only uses a two-stage stacked network.The network is tested on standard MPII dataset,the results are analyzed and verified.Experiments prove that the network effectively improves the accuracy.