2D Human Pose Estimation Based On Lightweight Networks

Human pose estimation is one of the important fundamental tasks of computer vision and the basis for the implementation of algorithms for action recognition,behavior recognition,and human-computer interaction.The current mainstream human pose estimation algorithms are highly complex,computationally intensive,and cannot run on resource-constrained devices such as the edge,which seriously limits the popularization of this technology.To address these problems,a lightweight human pose estimation network Mobile Pose Net based on refined module design and a lightweight human pose estimation network Ghost Pose Net based on knowledge distillation are proposed in this paper.The experimental results on the mainstream human pose datasets MPII and MSCOCO show that the proposed method achieves comparable performance with large networks with only 1.5M number of parameters and0.60 GFLOPs computation,respectively.The main work is as follows:1.A lightweight human pose estimation model based on refinement module is proposed.Unlike existing human pose estimation methods that use large convolutional kernels and design complex network structures to focus on accuracy,this paper proposes a lightweight upsampling module and designs a lightweight human pose estimation model with simple structure and low number of parameters based on this module.Among them,the lightweight upsampling module mainly includes three parts: point convolution,depth deconvolution and channel attention.Firstly,the input raw features are mapped to the high-dimensional space by point convolution,then the depth transpose convolution is used to improve the resolution of the input features,then the point convolution is used to compress the high-dimensional features back to the low-dimensional space,and finally the channel attention mechanism is used to encode the feature information of different channels to achieve the information fusion between different channels.The lightweight human pose estimation model adopts an encoder-decoder structure,using the first 14 layers of Mobile Net V3 as the body of the encoder,using the proposed lightweight upsampling module as the decoder of the model,and finally using 1 × 1convolution for the regression of key points.The results show that Mobile Pose Net achieves performance equivalent to that of a large human pose estimation network with a parametric weight of 1.5M.2.A lightweight human pose estimation model based on knowledge distillation is proposed.In order to further improve the efficiency and accuracy of the model operation,unlike existing knowledge distillation methods for human pose estimation that use networks with similar structure to the teacher network and small complexity differences as the student network,this paper proposes an online optimal knowledge distillation strategy based on joints and designs to use Ghost Pose Net,a lightweight human pose estimation network with low calculations and simple network structure,as the student network and HRNet with high network complexity as the teacher network.In order to solve the problem that the student network is difficult to optimize due to the large difference in teacher-student network complexity,the teacher-student network is constrained using MSE loss in the process of model training,and then the nodal bias is generated using the labels and the output of the student network,allowing the student network to focus on the output of the teacher network.Experimental results on the MSCOCO dataset show that Ghost Pose Net achieves performance equivalent to that of a large human pose estimation network with 0.6 GFLOPs.3.A pose matching system based on a lightweight human pose estimation model is developed.In order to solve the problems of time-consuming and inefficient manual pose correction,expensive professional sports equipment,etc.,and to realize fitness pose correction anytime and anywhere,this chapter develops the human pose matching system.The pose matching system adopts Windows 10 operating system and is developed based on Py Qt5 and Pytorch framework.It mainly contains four modules: user login and registration module,database module,human pose detection module and human pose matching.Among them,the human body detection module adopts the lightweight human posture estimation model proposed in this paper,which makes the system run with lower CPU and memory usage.In addition,the human pose matching mainly contains absolute pose matching and relative pose matching functions,which realize the matching of pose at different scales.