Multi person pose estimation based on improved openpose model

Abstract In order to solve the problem of multi person pose estimation, a human body pose estimation model framework based on openpose from American Carnegie Mellon University is presented, which uses the COCO (Microsoft common objects in context) data set to augment the data, optimize the network structure and retrain, adjust the learning rate properly and the convergence speed in the early stage of the model, finally using the model to carry out research on human key point detection and visualizing analysis in a multi-person complex scene. It is found that the improved openpose model adopts a bottom to up detection strategy to avoid the influence of the number of human bodies on real-time performance, and has higher key point detection efficiency than the top to down method. The study provides a real-time position feedback scheme for human pose estimation in complex scenes.