Analysis And Research On Human Pose Estimation And Evaluation Method Of CPR Using Multi-view Video

Cardiopulmonary resuscitation(CPR)is the most basic method to save the lives of people in respiratory and cardiac arrest.At present,there are few rescuers in society who master CPR rescue skills,and the efficiency of offline training is low,so the use of computer systems "self-service" type to complete CPR training is very meaningful.To achieve the objective,we designed and implemented a method for human pose estimation and CPR pose evaluation.In the human pose recognition task,a two-stage approach is used in this paper.Firstly,a 2D convolutional backbone network is used to obtain the heatmap of the human keypoints of the input image,and then the 2D body keypoint coordinates are derived by heatmaps using regression method.The 2D body keypoint coordinates are then reconstructed into 3D keypoint coordinates using triangulation method with multi-view camera parameters.Finally,the backbone network is improved to address the shortcomings of the traditional triangulation method,and the output weight parameters control the contribution of 2D keypoints from different views in the triangulation process.For the CPR action evaluation task,we defined the features to measure whether the action is qualified or not using the human 3D keypoint coordinates,and converted the CPR action evaluation problem into a time series binary classification problem.Then,the MLSTM-FCN network for time series classification is built based on the LSTM-FCN model to predict the action evaluation results.In addition,to address the extreme lack of CPR maneuver datasets in the community,we also created datasets for CPR action.The dataset contains 402 synchronized action videos of CPR practitioners from 5 viewpoints,where each video lasts approximately 40 s and includes 1000 frames of images.Each subject’s action video has the corresponding 3D human skeleton obtained by the human pose estimation model.Finally,we invited experts to manually annotate the dataset for qualified actions as labels for the data to be used in training human action evaluation model.The experimental results show that the CPR human pose estimation and CPR action evaluation method proposed in this paper can accurately predict the 3D human skeleton points and accurately determine whether the subjects are qualified for CPR action.