Human Pose Estimation Based On Deep Learnin

Human pose estimation task aims to locate the position of human key points from images or videos.This task can be used as the basis of other computer vision tasks.In recent years,human pose estimation has attracted more and more attention,and has been widely used in human-computer interaction,computer simulation system and other real scenes.The current algorithm based on deep learning overcomes the disadvantage of strong subjectivity of manual feature extraction,and has strong feature learning ability.However,the problems of occlusion and insufficient training data still restrict the improvement of the detection accuracy of the algorithm,and the complex network structure will lead to too slow detection speed.In view of the above problems,this paper puts forward three solutions:(1)A human pose estimation algorithm based on dynamic human perception convolution is proposed.According to the complex characteristics of human body,deformable convolution is introduced to enable the network to adaptively adjust the receptive field and enhance the ability of the backbone network to aggregate multi-scale spatial information.The conditionally parameterized convolution is used to design the keypoint detection module,which avoids the problem that the detection speed of the algorithm depends on the number of human instances.Accurate keypoint coordinates are obtained through the keypoint alignment module based on regression.Experimental results on a public dataset show that the algorithm can achieve a good balance between detection accuracy and speed.(2)An end-to-end human pose estimation algorithm based on bidirectional fusion feature pyramid is designed.A two-way fusion feature pyramid network is established,which realizes efficient feature extraction through two-way cross-scale connection and weighted feature fusion.Taking the minimum circumscribed rectangle of human keypoints as the bounding box for training,the idea of target detection is applied to human pose estimation.Experimental results on a public dataset show that the algorithm can achieve efficient human keypoint detection.(3)A human pose estimation algorithm based on swin transformer backbone and position encoder is explored.Swin transformer is introduced as the backbone network,and the network structure is optimized according to the characteristics of human pose estimation task.By compressing the original image information into a compact position sequence of keypoints,the human pose estimation task is transformed into a coding task.The keypoint dependencies are obtained by calculating the attention score,and the final keypoint position is predicted.Experimental results on a public dataset show that the algorithm can effectively predict the location of occluded human keypoints.