A Systematic Research Of Human Pose Estimation Based On Deep Learning

Human pose estimation is a basic research direction in the field of computer vision and has important research value.It is also a key technology in derivative applications such as pose transfer and action recognition.In recent years,the research on human pose estimation has ushered in the wave of deep learning.Relying on the powerful fitting ability of neural networks,the research on human pose estimation has developed rapidly.Based on the current research status,this thesis conducts further research on human pose estimation aiming at improving the accuracy and achieving lightweight models.In this research,we aim at improving the accuracy.In order to extract the multi-scale features that are very important for human pose estimation,the existing human pose estimation network often includes upsample and downsample processes,in which the downsample process is a process of compressing and refining feature units,which may lead to the loss of some feature information.For a high-precision vision task such as human pose estimation,the loss of features during downsampling may directly affect the upper bound of pose estimation accuracy.Although some existing methods have noted this problem,they have not properly addressed it.Aiming at this problem,this thesis proposed a pose estimation method based on intrinsic feature preserving,which optimizes the feature loss problem in the downsample process from multiple perspectives such as model,strategy,and keypoint representation.In the experiments,the proposed method achieves the new SOTA(state-of-the-art)performance on the MPII test set.In this research,we aim at achieving lightweight models.According to the analysis of the current lightweight human pose estimation research status,our work in this thesis is based on two core ideas.The first is to add efficient optimization modules and training strategies with only a small or no increase in inference parameters and computational complexity.The second is to greatly reduce the number of parameters and computational complexity under the premise of ensuring the least reduction of accuracy.Based on the above ideas,this thesis proposed four core points of the research on lightweight human pose estimation,namely model structure,training strategy,data enhancement,and keypoint representation.Based on the optimization of these four points,a more efficient lightweight human pose estimation method is proposed.In the experiments,the proposed method shows excellent accuracy performance on the MPII dataset,and on the COCO test set,it achieves the new lightweight human pose estimation SOTA performance.Furthermore,based on the proposed lightweight human pose estimation method,in this thesis,an Android-based human pose estimation prototype system is designed and developed.The system uses the mobile phone camera to collect real-time images as the input and uses NCNN as the model deployment framework.Based on the top-down human pose estimation scheme,this prototype system uses YOLO-Nano as the human body detection model to obtain the human body region subgraph,and then uses the lightweight human pose estimation method proposed in this thesis to estimate the human pose.This prototype system provides a general solution for the application of the human pose estimation algorithm on the mobile terminal.