Research On 2D Human Pose Estimation Based On High-resolution Networ

Human pose estimation is a technique for predicting human pose by detecting key points of the human body in images and predicting human pose based on the connection relationship of limb parts,which is widely used in human modeling,human-computer interaction and medical rehabilitation.In recent years,human pose estimation based on high-resolution networks has shown excellent performance compared to other models,but its high number of parameters and high computational complexity limit its deployment in practical applications.In addition,factors such as human self-occlusion,mutual occlusion and multiple numbers that occur in multi-crowded states make the performance of multi-human pose estimation by high-resolution variability networks also perform poorly.To solve the above problems,this paper focuses on efficient network design based on high-resolution networks and improving the performance of multicrowded human pose estimation as the key research objectives.The details of the research are as follows:(1)An efficient Lite-Higher HRNet based on high-resolution network for human pose estimation is proposed.The simple and efficient network model is designed by removing the redundant switching units and the base residual module in the highresolution network through multiple reduction experiments;the base unit of the model is replaced with an efficient residual bottleneck module,in which a convolutional attention module is introduced to improve the model accuracy,achieving the goal of further compressing the number of parameters and computational complexity of the model.The results of training and testing on the MSCOCO dataset show that Lite-Higher HRNet has less number of parameters and computational complexity than other large models,and achieves the design goal of lightweight model,in which the performance is partially lost,but still maintained at 61.3.(2)SRE-UPose for unbiased human pose estimation based on regularized embedding grouping is proposed.The unbiased data processing method is proposed to solve the problem of quantization error arising from data pre-processing;Then,the key point grouping algorithm with scale regularized embedding vectors is designed to improve the accuracy of key point grouping in multi-congestion states;Finally,the key point coding and decoding method is improved,and the weight adaptive heat map regression method is proposed to solve the problem of background sample overfitting during the training process.The results of training and testing on the Crowd Pose dataset show that the performance is improved compared with other models and performs optimally with the highest congestion metric.