Research And Application Of Human Key Point Detection In Pose Estimtion

Human pose estimation is one of the important directions in artificial intelligence,it is also a hot research topic in the field of computer vision in recent years,and has been widely applied in security systems,somatosensory technology and human-computer interaction.Researchers in human pose estimation have found that the task of human pose estimation is full of challenges,including complex background of natural pictures,the disordered position of persons,severe occlusion of human key points,and the difficulty in defining the relationship between the key points.Currently researches on pose estimation are mainly from the detection of skeleton joints.However,most of these methods focus on single person,or making the task of multi-person pose estimation into human object detection and joints detection of pose estimation steps.This is,however,ignores the global information and context information of the image,and leading to the low speed of human estimation,and the poor performance of multi-human body estimation.This thesis firstly conducts a literature survey,and the fully investigation of public datasets and related technologies in the field of human pose estimation.Secondly,based on the top-down human pose estimation scheme,this thesis combines the current excellent object detection network Yolov3 respectively with the cascade pyramid deep convolution network and the stacking hourglass network to realize the multi-person pose estimation algorithm.Experiments on COCO dataset prove that the model convergence speed and accuracy of the cascade pyramid deep convolution network based joint detection method are higher than the stacked hourglass network based detection method.The thesis then proposes a multi-person human pose estimation algorithm by compromising associative embedding and cascade pyramid deep convolution network.The algorithm simultaneously accomplishes the tasks of person detection and human pose estimation in the one-stage and end-to-end neural network structure.The network extracts local and global feature information of the image based on the cascade pyramid deep neural network,and measures the joints information of multiple people by the associative embedding.According to the experiments on COCO human skeleton joints dataset,the end-to-end multi-person pose estimation algorithm can effectively detect multi-person skeleton key points,that has high accuracy and is robust to occlusion and noise.