Human Key Point Detection Algorithm Based on Improved OpenPose
Human key point detection is widely used in the field of human posture estimation.To solve the problems of slow de-tection speed and inability to achieve global optimal matching of key points in multi-person scenes,this paper proposes a human key point detection algorithm based on improved OpenPose.Firstly,the ordinary 3D convolution used in the classic OpenPose pre-feature extraction network is replaced by DSC to reduce the scale of model parameters and improve the detection speed.Then,aiming at the coordinate label values corresponding to the key point coordinate regression branch(PCM),a labeling strategy based on Gaussian kernel is proposed,which makes the network training process more robust.Finally,based on Hungarian algorithm to do bipartite matching,the global optimal matching of key points in multiplayer scene is realized.In this paper,the algorithm is eval-uated on the COCO2017 data set.In the ablation test,the detection frame rate FPS reaches 34,which is 36%higher than the clas-sic OpenPose,and the corresponding AP50,AP75 and AP90 indexes reach 92.5,81.4 and 70.8 respectively,which are 8.4%,9.1%and 8.9%higher than other key point detection schemes.
human keypoint detectionOpenPoseconvolutional neural networkmatch of bipartite graph
万方数据
维普
