Human Pose Estimation Based On Lightweight Network At Low-Resolution

Human pose estimation is an important research direction in the field of computer vision,which is widely used in human-computer interaction,behavior analysis,and intelligent surveillance.Although existing human pose estimation algorithms can obtain high accuracy through complex network structures,they are difficult to be widely used in people’s daily life due to their high number of parameters and complexity of operations,as well as their poor performance for low-resolution images that are common in practical applications.In this thesis,we have conducted an in-depth study to address the above problems using techniques related to network lightweighting and super-resolution reconstruction,and the main work is listed as follows:(1)A lightweight network for human pose estimation is proposed.Two lightweight modules,SG-neck(S-Ghost Bottleneck)and SD-block(Sandglass Dual Channel Attention Basicblock),are designed to replace the standard convolution operation to ensure the accuracy of the model while reducing the number of parameters and the complexity of the operation.In the SG-neck module,the S-Ghost module is built to replace the 1×1 standard convolution by combining the channel mixing and Ghost module to enhance the interaction between channels,and the number of channels is increased to extract rich feature information;in the SD-block module,the Sandglass module is used to replace the 3×3 standard convolution and embed the DCA(Dual Channel Attention)module to focus on the target features to enhance the expressiveness of the model.Experimental results on MPII and COCO datasets show that the proposed network achieves a better balance between accuracy and model complexity compared with many different types of human pose estimation algorithms.(2)A human pose estimation network for low-resolution images is proposed.First,the super-resolution network(LapSRN)is used as the pre-processing module of the human pose estimation network,and the stepwise reconstruction features of LapSRN are incorporated into the human pose estimation network to capture more detailed information of the human body;second,the Gaussian heat map is transformed using the β-Soft-Argmax post-processing function to alleviate the quantization error problem caused by the key point coding and decoding method;finally,the LapSRN network is jointly trained by the compound loss function to reconstruct images that are favorable for human pose estimation.The experimental results on the COCO dataset show that the proposed network in this thesis can significantly improve the accuracy of human pose estimation under low-resolution images compared with other human pose estimation algorithms.(3)A human pose estimation system in low-resolution scenes is designed,which enables uploading,pose estimation and saving functions for low-resolution images.The system,combined with the model proposed in Chapter 4 of this thesis,enables users to achieve the prediction of human posture in low-resolution scenes intuitively and conveniently,and has good practical application value.