Research On Human Pose Estimation Algorithm Based On Deep Learning

Human pose estimation is one of the current challenging and popular research directions,which aims to detect the human keypoints of each human body from pictures or videos to describe the human skeleton.With the mature development of human pose estimation,human pose estimation is widely used in intelligent surveillance,virtual reality,motion analysis and so on.In this paper,we study and improve the human pose estimation algorithm based on deep learning,and the main research contents are as follows:(1)A human pose estimation algorithm based on dual-attention and multi-scale fusion is proposed.In the human pose estimation task,features extracted by convolutional neural networks contain information of different levels of importance.This paper uses a dual-attention mechanism to perform feature enhancement on the features extracted by the backbone network to obtain richer and more important detailed information.And the down-sampling operation in the feature extraction process will lose a lot of spatial information.This paper increases the feature perceptual field by stacking atrous convolution with different expansion rates and keeps the resolution size of the features constant,so that the network model obtains multi-scale feature information of human keypoints.And the obtained features of different scales are fused to improve the localization accuracy of human keypoints and enhance the network performance.The experimental results show that the detection accuracy of human keypoints can be improved by enhancing the features and performing multi-scale fusion.(2)A human pose estimation algorithm based on contextual information aggregation is proposed.In this paper,a self-attention mechanism is introduced in the feature extraction stage to capture the contextual information of the feature map at a distance,obtain the dependencies between remote pixel points to obtain global constraints,extract global contextual information and perform aggregation.For the multi-level features extracted by the backbone network,a multi-level branching feature fusion mechanism is designed to combine features from different stages,expand the feature maps step by step,and recover the feature maps resolution to output the predicted keypoints heatmaps.In the process of getting the predicted keypoints heatmaps to the human keypoints coordinates,distribution-aware coordinate decoding is used to significantly improve the accuracy of coordinate regression.The experimental results show that the algorithm can detect human keypoints more accurately and further improve the network model performance.(3)A human pose estimation algorithm based on graph encoding optimization is proposed.In the human pose estimation task,there are certain connection relationships between the human body joints.In this paper,we model the connection relationship between adjacent human body joints,establish the skeleton diagram of human body joints as prior knowledge based on the structural model of human body graph,and use the connection relationship between adjacent joints of human body to make the network learn the joint connection information to assist the detection of human keypoints.The final experimental results show that the method can achieve better algorithm performance by introducing prior knowledge to optimize the human pose estimation algorithm.